feat(pairs): attiva ETH/BTC 15m flat-skip in PORT06 (BLEND, mezza size)

Origine: gioco "Blind Traders" (100 agenti ciechi su BTC/ETH anonimizzati) ->
vincitore = spread ETH/BTC reversion a 15m. Testato sul serio col gate PORT06:
non duplicato (corr 1h vs 15m = 0.37), robusto (16/16 celle Sharpe>1), edge NON
artefatto delle candele flat ETH 15m (filtrandole resta l'83% dello Sharpe).

Percorso live costruito e validato:
- pairs_research.pairs_sim_flat: engine generalizzato con exit LIVE-REALIZABLE
  (arma exit_ready, esce alla 1a barra pulita); regression-lock a pairs_sim.
- PairsWorker: flat_skip + exit_ready + rilevamento flat da OHLC (1h byte-exact).
- runner: fetch diretto dei timeframe sub-orari + override position_size per-sleeve.
- validate_worker_pairs: replay worker == backtest a 15m (8452 vs 8453 trade).
- _defs/build_everything: sleeve PR_ETHBTC_15M (mezza size, pos 0.10) -> PORT06
  FULL 6.43->7.20, OOS 8.58->9.66, DD giu'. Rischio bilanciato col 1h.
- smoke live: Cerbero serve candele 15m fresche; worker ticca.

Diari docs/diary/2026-06-09-*. Caveat slippage: mezza size = blend-tilt prudente.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Adriano Dal Pastro
2026-06-09 11:48:15 +00:00
parent 5d45f4ef6e
commit d25d897fd1
20 changed files with 1727 additions and 60 deletions
+3
View File
@@ -32,3 +32,6 @@ data/regime/dispersion_features.parquet
# storico catena opzioni importato da cerbero-bite (rigenerabile: options_fetcher.py) # storico catena opzioni importato da cerbero-bite (rigenerabile: options_fetcher.py)
data/options/ data/options/
data/_reset_backup/ data/_reset_backup/
# game artifacts (log/json di scripts/games e gate)
data/games/
@@ -0,0 +1,56 @@
# 2026-06-09 — Gioco "Blind Traders": 100 agenti ciechi
## Setup
100 agenti LLM (haiku) ricevono due serie anonime **X** e **Y** — in realta'
**BTC** e **ETH** 1h/15m/5m, mai etichettate — e devono proporre UNA regola che
"anticipi" i movimenti per un PnL netto positivo (fee 0.10% RT) con **>=10
trade/mese**. Non sanno cosa siano i dati. L'orchestratore (engine deterministico)
valuta ogni strategia, assegna un punteggio su **PNL + %win**, da' **90 epoche di
elaborazione** (hill-climb dei parametri) e **ogni 10 epoche blocca il 10% meno
profittevole** -> restano i **10 piu' profittevoli**.
Infrastruttura in `scripts/games/`:
- `engine.py` — dati anonimizzati, 6 famiglie segnale (zscore/breakout/ma_cross/
rsi/momentum/pairs), backtester causale fee-aware, scoring (>=10 tpm o squalifica).
- `agent_brief.py` — digest ANONIMO (stat aggregate + finestra normalizzata) + menu.
- `arena.py` — torneo a **3 finestre**: TRAIN (hill-climb), VALID (cull+rank
dell'orchestratore), TEST (OOS puro, mai ottimizzato). Anti-overfit.
- `run_game.py` — carica le 100 spec degli agenti e lancia il torneo.
## Risultato emergente
I 100 agenti ciechi, leggendo SOLO le statistiche anonime (autocorrelazione
negativa, "after_big_move_continues_pct" ~30-40% => le mosse estreme rientrano),
hanno **riscoperto da soli che il mercato e' mean-reverting**: 100/100 reversion,
67 hanno scelto il detector pairs, 30 zscore. Esattamente la lezione storica del
progetto (edge = reversione; pairs ETH/BTC il piu' robusto) — senza sapere che
fosse crypto.
## Classifica finale (top 10) — tutti PAIRS su 15m
Vincitore **agente #91** (15m, pairs market-neutral sul log-ratio X/Y):
- TEST/OOS puro: **PnL +3126%**, **win 77%**, **108.9 trade/mese**, **Sharpe 20.3**
- Full-period: PnL +8052%, win 70%, 94 tpm, Sharpe 12.2 (9604 trade)
- params: lookback 66, entry 1.67σ, exit 1.0σ, max_bars 35
- ipotesi (cieca): "Y altamente reversivo, X/Y log-ratio strong mean-reversion
(-0.43 autocorr), bassa correlazione cross-asset -> pairs market-neutral".
Tutti i 10 finalisti: pairs 15m, TEST Sharpe medio 19.9, tpm 66-109 (>>10).
## Caveat onesti
- Numeri OOS ottimistici: PnL additivo a notional fisso, **niente slippage sulle 2
gambe**, finestra OOS calma, 15m molti trade. Coerente col caveat PR01 del
progetto (Sharpe reale atteso ~4-5, non 20). Il valore del gioco e' il **metodo**
(scoperta cieca + selezione anti-overfit), non il livello assoluto di Sharpe.
- La convergenza su pairs conferma robustezza ma riduce la diversita': i 10 finalisti
sono varianti della stessa idea (ETH/BTC spread). Per un portafoglio servirebbe
diversificare (gia' fatto altrove: fade + honest + shape).
## Re-run "sobrio" con slippage (0.05%/lato)
`GAME_SLIP=0.0005` -> i pairs pagano +0.20% RT extra (4 lati). Lo slippage spinge
l'ottimizzatore verso **meno churn**: tpm dei finalisti 66-109 -> **40-47**, Sharpe
top-10 ~20 -> ~13.5. Vincitore **#43** (15m pairs): TEST PnL **+2091%**, win 77%,
**46.9 tpm**, Sharpe **15.6**. La gerarchia (pairs 15m domina) e la robustezza
reggono lo stress; lo Sharpe reale atteso resta ~4-5 (OOS calmo + PnL additivo).
Log: `data/games/game_slip.log`.
Artefatti: `data/games/tournament_result.json`, `data/games/specs/agent_*.json`,
`engine.set_slippage()` (env `GAME_SLIP`).
@@ -0,0 +1,82 @@
# 2026-06-09 — Percorso live 15m per ETH/BTC pairs: COSTRUITO e VALIDATO
Seguito di `2026-06-09-pairs15m-port06-gate.md` (il gate passa, edge reale e non
artefatto flat). Qui si costruisce e VALIDA l'infrastruttura per eseguire il pairs
ETH/BTC a 15m con flat-skip, alla pari del backtest (disciplina validate_worker_pairs).
## 1. Engine canonico (regression-locked)
`scripts/analysis/pairs_research.py`: aggiunti `aligned_ohlc`, `is_flat_ohlc`,
`pairs_sim_flat(..., flat_skip, scan_buffer)`. Regola di uscita **LIVE-REALIZABLE**:
la condizione (|z|<=z_exit O bars>=max_bars) ARMA `exit_ready`; si esce al CLOSE della
PRIMA barra PULITA successiva (mai a un prezzo passato come faceva il prototipo push-back).
- **Regression-lock**: `pairs_sim_flat(flat_skip=False)` == `pairs_sim` ESATTO
(ETH/BTC 1h 1756 trade, 15m 9388 trade, ret/dd/sharpe identici al bit).
## 2. PairsWorker esteso (retrocompatibile)
`src/live/pairs_worker.py`: param `flat_skip`, stato `exit_ready` (persistito), tick
ora fa merge OHLC e rileva le candele flat (O=H=L=C in UNA gamba). Entry saltato su barra
stale; uscita con la stessa regola exit_ready dell'engine. **Default off = comportamento
1h storico invariato** (se mancano le colonne OHLC, flat=False).
## 3. Runner: fetch sub-orario (inerte finche' non c'e' uno sleeve 15m)
`src/portfolio/runner.py`: `_SUBHOURLY={5m,15m,30m}`, `_LOOKBACK_DAYS` esteso; il loop
fetcha DIRETTO da Cerbero i timeframe sub-orari per (asset,tf) (non resamplabili dal 1h) e
un router `_series_for` instrada la serie giusta a ogni worker. Zero impatto sul live
attuale: nessuno sleeve e' 15m → `subhourly_needs` vuoto → ramo morto.
## 4. VALIDAZIONE (validate_worker_pairs.py) — TUTTO OK
Replay bar-per-bar del worker == backtest:
| caso | worker | backtest | match |
|---|---|---|---|
| ETH/BTC 1h | 1756 trd, cap 2.886.616 | 1756, 2.886.616 | **OK esatto** |
| BTC/LTC 1h | 599 trd, cap 16.861 | 599, 16.861 | **OK esatto** |
| **ETH/BTC 15m-flat** | **8452 trd** | **8453 trd** (cap entro 0.15%) | **OK** |
(1 trade di differenza = posizione finale aperta non chiusa nel replay, atteso.)
## 5. Gate finale (engine == worker) — PROMOSSO
`pairs15m_gate_final.py` (corr 1h vs 15m = 0.372, 3201 ingressi flat saltati):
| variante ETH/BTC | FULL Sh | FULL DD | OOS Sh | OOS DD |
|---|---|---|---|---|
| baseline 1h | 6.43 | 3.96 | 8.58 | 1.36 |
| **SWAP 15m-flat** | 7.31 | 3.55 | **9.95** | **1.26** |
| **BLEND 1h+15m** | 7.03 | 3.66 | 9.57 | 1.24 |
Entrambi PROMOSSI (a fee backtest). Caveat slippage del gate precedente invariato → il
BLEND e' la forma raccomandata (meta' allocazione sul 1h pulito, slippage-robusto).
## Stato e attivazione (NON fatta — decisione di deploy)
Tutto il PERCORSO e' pronto e validato, ma il 15m **non e' attivo nel portafoglio live**:
attivarlo cambia il trading reale e va deciso esplicitamente. Per accenderlo:
1. `_defs.py`: aggiungere SleeveSpec pairs ETH/BTC a 15m (tf="15m",
params={n:66,z_in:1.674,z_exit:1.0,max_bars:35,flat_skip:True}) — come SWAP della 1h o
come 2a sleeve (BLEND) sotto il cap PAIRS.
2. `report_families.build_everything` / `sleeves`: l'equity del nuovo sleeve dal
`pairs_sim_flat(tf=15m, flat_skip=True)` (per parita' backtest==report).
3. Shadow smoke su testnet (come `live_smoke_pairs.py`) prima del paper reale.
4. `deploy.sh` (bump+rebuild) — il runner gia' fetcha 15m e passa flat_skip via spec.params.
Test suite: nessuna regressione (1h byte-exact). Artefatti: pairs_research.py,
pairs_worker.py, runner.py, validate_worker_pairs.py, pairs15m_gate_final.py.
## ATTIVAZIONE IN REALE (2026-06-09) — BLEND, mezza size
Deciso: BLEND (sleeve 15m ACCANTO al 1h, non swap). Implementato:
- `_defs.py`: SleeveSpec `PR_ETHBTC_15M` (tf=15m, flat_skip, params.position_size=0.10
= meta' del family PAIRS 0.20) in PAIRS -> entra in PORT04/05/06.
- `report_families.build_everything`: equity da `pairs_sim_flat(tf=15m, flat_skip=True, pos=0.075)`
(mezza size, == intento live) con sid PR_ETHBTC_15M.
- `runner.pos_for_spec`: override PER-SLEEVE (params.position_size) > famiglia > globale.
- **Mezza size perche'** a peso pieno il 15m pesava il 25.8% del rischio PORT06 (vs 9.5% del
1h): dimezzato -> 11.5% vs 10.6%, bilanciato. Disciplina come la cap SHAPE; rispetta il
caveat slippage (il 15m non domina il book).
**PORT06 col BLEND (mezza size)**: FULL Sharpe **6.43->7.20** DD **3.96->3.68**,
OOS Sharpe **8.58->9.66** DD **1.36->1.31**. Migliora tutto.
**Smoke live 15m** (`pairs15m_live_smoke.py`): Cerbero serve candele 15m FRESCHE per
ETH e BTC (ultima barra 0 min fa, flat live 2-3%), worker flat-skip ticca OK. Esecuzione
reale a 2 gambe gia' coperta da `live_pairs_smoke.py` (livello strumento, tf-indipendente).
**Regression-lock aggiornati** (miglioria attesa, non regressione): test_definitions
(17->18 sleeve), test_backtest_parity_cap (FULL 6.47->7.20, OOS 8.82->9.66). Suite verde.
Live: il runner fetcha 15m diretto, costruisce il PairsWorker(flat_skip) col pos 0.10,
e lo esegue reale a 2 gambe (pairs_enabled). Attivazione via deploy (bump+rebuild).
@@ -0,0 +1,89 @@
# 2026-06-09 — ETH/BTC pairs a 15m: gate PORT06 (dal gioco Blind Traders)
## Origine
Il gioco "Blind Traders" (100 agenti ciechi) ha eletto come vincitore una variante
ETH/BTC pairs su **15m** (config #43: n=66 z_in=1.67 z_exit=1.0 max_bars=35). Domanda:
e' un vero miglioramento o un duplicato piu' veloce della sleeve PR01 ETH/BTC gia'
deployata a 1h? Testato sul serio con l'engine di PRODUZIONE `pairs_sim` + gate PORT06.
Script: `scripts/analysis/pairs15m_port06_gate.py`.
## Risultati
- **Parita' OK** (corr 1.00000): l'harness riproduce esattamente il sleeve canonico
PR_ETHBTC → gate affidabile.
- **CORRELAZIONE 1h vs 15m = 0.349** (rendimenti giornalieri). **SMENTISCE la mia
ipotesi iniziale "duplicato ridondante"**: a 15m cattura eventi di reversione DIVERSI
→ e' un diversificatore reale, non una doppia scommessa sullo stesso spread.
- **Robustezza 15m**: griglia n×z_in → **16/16 celle Sharpe>1** (9-12), plateau non picco.
Non e' un punto overfit del gioco.
- **Standalone**: 15m fa 9388 trade (vs 1756 a 1h), Sharpe 11.7 (vs 4.36), DD 54% (vs 48%),
8/9 anni+ . (Le % FULL sono esplose dal compounding pos0.15·lev3 su 9k trade → guardare
Sharpe/DD/anni, non il livello %.)
## Gate PORT06 (pos0.15 lev3 canonico, OOS da 2024-10-12)
| variante ETH/BTC | FULL Sh | FULL DD | OOS Sh | OOS DD |
|---|---|---|---|---|
| **baseline 1h** | 6.43 | 3.96 | 8.58 | 1.36 |
| **SWAP 15m** | 7.64 | 3.49 | **10.39** | **1.26** |
| **BLEND 1h+15m** | 7.30 | 3.63 | 9.95 | 1.24 |
A fee di backtest (0.20% RT/coppia) **entrambe PROMOSSE**: Sharpe su e DD giu' ovunque.
## Stress slippage a livello PORT06 (il vero rischio: 15m = 5× i trade)
| fee_rt | RT/coppia | PORT06 FULL Sh | FULL DD | OOS Sh | OOS DD | std Sh | std oDD |
|---|---|---|---|---|---|---|---|
| baseline 1h | 0.20% | 6.43 | 3.96 | 8.58 | 1.36 | 4.36 | 16% |
| 15m | 0.20% | 7.64 | 3.49 | 10.39 | 1.26 | 11.7 | 13% |
| 15m | 0.40% | 7.04 | 4.08 | 9.78 | 1.45 | 8.5 | 27% |
| 15m | 0.60% | 6.43 | 4.67 | 9.15 | 1.66 | 5.3 | 47% |
**Degradazione graziosa ma reale**: il vantaggio di **Sharpe** sopravvive fino a slippage
pessimista (OOS 9.15 > 8.58 anche a 0.60%), ma il vantaggio di **DD si perde gia' a 0.40%**
(FULL DD 4.08 > 3.96 baseline; standalone oDD esplode 13→27→47%). La regola del progetto
("ri-gateare ogni filtro quando cambiano i costi") qui taglia: la frequenza 5× rende la
sleeve slippage-sensitive.
## Verdetto
- **NON un duplicato** (corr 0.35) e **NON overfit** (16/16 robusto) → la mia liquidazione
iniziale era SBAGLIATA, lo dico chiaro.
- **Passa il gate a fee di backtest, marginale sotto slippage**: migliora Sharpe sempre, ma
sotto slippage realistico (≥0.40% RT) peggiora leggermente il DD di portafoglio.
- **Due rischi di produzione NON ancora quantificati**: (a) qualita' dati ETH 15m (14-30%/anno
candele flat O=H=L=C → fill non eseguibili che gonfierebbero il backtest), (b) fill/liquidita'
reale a 2 gambe a 15m (5× ordini). Il worker pairs e' validato a 1h, non a 15m.
**Raccomandazione**: NON swap diretto in live. Candidato promettente → percorso forward:
preferire il **BLEND 1h+15m** (tiene il DD pulito del 1h e raccoglie il rendimento
decorrelato del 15m) **dopo** un check sull'impatto delle candele flat 15m sui pairs.
Allineato a come il progetto tratta FR01 (robusto ma non deployato finche' non domina pulito).
Resta come record di ricerca; deploy solo se il check flat-candle e' pulito.
## CHECK FLAT-CANDLE (pairs15m_flatcheck.py) — PULITO
Rischio: ETH 15m ha molte candele flat (O=H=L=C) → close stale che gonfia z-score →
reversione FINTA non eseguibile. Test:
- **Prevalenza**: ETH 15m **16.4% medio** (fino 30% nel 2022); BTC 15m solo 3.5%. Reale.
- **Fill toccati**: 12.9% degli entry e 15.2% degli exit cadono su una barra flat.
- **Test decisivo** (entry/exit SOLO su barre pulite, non-flat in entrambe le gambe):
rimuove 11.2% dei trade, **Sharpe trattenuto all'83%** (11.74→9.70; OOS Sharpe 18.4).
Se l'edge fosse un artefatto flat, filtrando crollerebbe → **NON crolla. NON e' artefatto.**
- **Gate PORT06 col 15m FLAT-FILTRATO** (corr 1h vs 15m-flat = 0.366, ancora decorrelato):
- SWAP 15m-flat: FULL 7.32/3.55, OOS **9.99/1.26** → PROMOSSO
- BLEND 1h+15m-flat: FULL 7.05/3.66, OOS **9.60/1.24** → PROMOSSO
## Conclusione (3 box su 4 puliti)
✅ NON duplicato (corr 0.35-0.37) ✅ robusto (16/16) ✅ NON artefatto flat (83% Sharpe)
⚠️ slippage-sensitive: a fee backtest passa pulito; a slippage ≥0.40% RT il vantaggio di
Sharpe regge ma il DD-edge si assottiglia. Il **BLEND** mitiga (meta' allocazione resta sul
1h pulito e slippage-robusto) → e' la forma deployabile.
## Realta' del deploy (perche' NON tocco ancora il live)
Il gate passa a livello BACKTEST. Ma il live NON puo' eseguire un sleeve 15m oggi:
- la live pairs gira SOLO a 1h (`PairsWorker`, validato da `validate_worker_pairs` a 1h);
il runner risampla a 1h/4h/1d, non gestisce un leg pairs a 15m.
- un BLEND richiede DUE sotto-sleeve ETH/BTC (1h + 15m) dentro il cap PAIRS, e il
**flat-skip va replicato nel worker live** (altrimenti il live tradera' le barre stale che
il backtest esclude → divergenza backtest-vs-live, la classe di bug che il progetto teme).
Editare `_defs.py` cambierebbe solo il backtest/report, NON il live → sarebbe ingannevole.
**Percorso deploy corretto** (da confermare): (1) estendere `PairsWorker`/runner al 15m +
flat-skip; (2) `validate_worker_pairs` a 15m (replay == backtest esatto); (3) aggiungere lo
sleeve 15m sotto il cap PAIRS; (4) shadow su testnet prima del paper. Finche' (1)-(2) non
sono fatti e validati, resta **record di ricerca PROMOSSO ma non live**.
+203
View File
@@ -0,0 +1,203 @@
"""Check candele FLAT (O=H=L=C, liquidita' zero) sui pairs ETH/BTC a 15m.
Rischio noto (CLAUDE.md): ETH 15m ha 14-30%/anno di candele flat per bassa liquidita'
del perpetuo. Su un pairs, un close stale gonfia lo z-score (l'altra gamba si muove,
questa e' ferma) -> segnale di "reversione" FINTO che rientra solo quando la gamba
stale si sblocca: profitto NON eseguibile dal vivo. Questo gonfierebbe il backtest 15m.
Test:
[1] prevalenza candele flat per anno (ETH 15m, BTC 15m).
[2] quanti trade del pairs 15m hanno ENTRY/EXIT su una candela flat (gamba stale).
[3] re-sim flat-aware: entry/exit SOLO su barre pulite (non-flat in ENTRAMBE le gambe)
-> quanto sopravvive l'edge? (parita': senza flat-skip == pairs_sim).
[4] gate PORT06 col 15m flat-filtrato vs baseline 1h.
uv run python scripts/analysis/pairs15m_flatcheck.py
"""
from __future__ import annotations
import sys
from pathlib import Path
import numpy as np
import pandas as pd
PROJECT_ROOT = Path(__file__).resolve().parents[2]
sys.path.insert(0, str(PROJECT_ROOT))
from src.data.downloader import load_data
from scripts.analysis.pairs_research import pairs_sim, OOS_FRAC, FEE_RT, LEV, POS, BARS_YEAR
from scripts.analysis.report_families import daily_from
from scripts.analysis.combine_portfolio import metrics, SPLIT, OOS_DATE
from scripts.analysis.pairs15m_port06_gate import port_metrics, eth_btc_daily, UNIV_1H, GAME_15M
from scripts.portfolios._defs import PORTFOLIOS
from src.portfolio.sleeves import all_sleeve_equities
def aligned2(a, b, tf="15m"):
"""Merge con OHLC di ENTRAMBE le gambe (serve per rilevare i flat su entrambe)."""
da = load_data(a, tf)[["timestamp", "open", "high", "low", "close"]].rename(
columns=lambda x: x + "_a" if x != "timestamp" else x)
db = load_data(b, tf)[["timestamp", "open", "high", "low", "close"]].rename(
columns=lambda x: x + "_b" if x != "timestamp" else x)
m = da.merge(db, on="timestamp", how="inner").reset_index(drop=True)
m["dt"] = pd.to_datetime(m["timestamp"], unit="ms", utc=True)
return m
def is_flat(o, h, l, c):
return (o == h) & (h == l) & (l == c)
def flat_prevalence(asset, tf="15m"):
d = load_data(asset, tf)
d = d.copy()
d["dt"] = pd.to_datetime(d["timestamp"], unit="ms", utc=True)
fl = is_flat(d["open"].values, d["high"].values, d["low"].values, d["close"].values)
d["flat"] = fl
by = d.groupby(d["dt"].dt.year)["flat"].mean() * 100
return by, fl.mean() * 100
def pairs_sim_flataware(a, b, tf="15m", n=66, z_in=1.674, z_exit=1.0, max_bars=35,
jump_max=0.08, fee_rt=FEE_RT, lev=LEV, pos=POS,
split_frac=0.0, skip_flat=True):
"""Come pairs_sim ma: entry/exit consentiti SOLO su barre pulite (se skip_flat).
Ritorna anche n_entry_flat / n_exit_flat (diagnostica, calcolata sempre)."""
m = aligned2(a, b, tf)
ca, cb = m["close_a"].values, m["close_b"].values
flat_a = is_flat(m["open_a"].values, m["high_a"].values, m["low_a"].values, ca)
flat_b = is_flat(m["open_b"].values, m["high_b"].values, m["low_b"].values, cb)
flat = flat_a | flat_b # barra "sporca" se una delle due gambe e' flat
r = np.log(ca / cb)
dr = np.abs(np.diff(r, prepend=r[0]))
ma = pd.Series(r).rolling(n).mean().values
sd = pd.Series(r).rolling(n).std().values
z = (r - ma) / np.where(sd == 0, np.nan, sd)
ts = m["dt"]; N = len(r)
split = int(N * split_frac)
fee = 2 * fee_rt * lev
cap = peak = 1000.0; dd = 0.0; last = -1
trades = wins = 0; rets = []; yearly = {}
eq_ts, eq_v = [], []
n_entry_flat = n_exit_flat = 0
for i in range(n + 1, N - 1):
if i < split or np.isnan(z[i]) or dr[i] > jump_max or i <= last:
continue
if z[i] <= -z_in:
d = 1
elif z[i] >= z_in:
d = -1
else:
continue
if flat[i]:
n_entry_flat += 1
if skip_flat:
continue # non si entra su una gamba stale
# exit: |z|<=z_exit o max_bars; se skip_flat, salta le barre flat come uscita
j = min(i + max_bars, N - 1)
for k in range(1, max_bars + 1):
jj = i + k
if jj >= N:
j = N - 1; break
if skip_flat and flat[jj]:
j = jj # avanza, non esce su barra stale
continue
if abs(z[jj]) <= z_exit:
j = jj; break
j = jj
if flat[j]:
n_exit_flat += 1
if skip_flat:
# spingi all'ultima barra pulita entro l'orizzonte
back = j
while back > i and flat[back]:
back -= 1
j = back if back > i else j
retA = (ca[j] - ca[i]) / ca[i]
retB = (cb[j] - cb[i]) / cb[i]
ret = (retA - retB) * d * lev - fee
cap = max(cap + cap * pos * ret, 10.0)
peak = max(peak, cap); dd = max(dd, (peak - cap) / peak)
trades += 1; wins += ret > 0; rets.append(ret * pos); last = j
eq_ts.append(ts.iloc[j]); eq_v.append(cap)
yearly[ts.iloc[i].year] = yearly.get(ts.iloc[i].year, 0.0) + ret * 100
yrs_span = (ts.iloc[-1] - ts.iloc[max(split, 0)]).days / 365.25 or 1
sharpe = 0.0
if len(rets) > 1 and np.std(rets) > 0:
sharpe = float(np.mean(rets) / np.std(rets) * np.sqrt(trades / yrs_span))
ret_tot = (cap / 1000 - 1) * 100
return dict(trades=trades, win=wins / trades * 100 if trades else 0, ret=ret_tot,
dd=dd * 100, sharpe=sharpe, yearly=yearly, eq_ts=eq_ts, eq_v=eq_v,
n_entry_flat=n_entry_flat, n_exit_flat=n_exit_flat)
def main():
print("=" * 100)
print(" CHECK FLAT-CANDLE — ETH/BTC pairs 15m (gate condizionato)")
print("=" * 100)
# [1] prevalenza
print("\n[1] Prevalenza candele flat (O=H=L=C) per anno, 15m:")
for asset in ("ETH", "BTC"):
by, tot = flat_prevalence(asset, "15m")
print(f" {asset}: media {tot:.1f}% | " +
" ".join(f"{y}:{v:.0f}%" for y, v in by.items()))
# [2] quanti trade toccano un flat (sim SENZA skip per diagnostica)
diag = pairs_sim_flataware("ETH", "BTC", **GAME_15M, skip_flat=False)
tr = diag["trades"]
print(f"\n[2] Trade 15m totali: {tr} | entry su barra flat: {diag['n_entry_flat']} "
f"({diag['n_entry_flat']/tr*100:.1f}%) | exit su barra flat: {diag['n_exit_flat']} "
f"({diag['n_exit_flat']/tr*100:.1f}%)")
# [3] parita' + edge filtrato
print("\n[3] Edge 15m: NO-skip (== pairs_sim) vs FLAT-AWARE (entry/exit solo barre pulite):")
# parita': flataware skip_flat=False deve ~== pairs_sim
base_ps = pairs_sim("ETH", "BTC", **GAME_15M, pos=POS, lev=LEV)
print(f" parita' pairs_sim : trd {base_ps['trades']:>5d} Sh {base_ps['sharpe']:.2f} "
f"DD {base_ps['dd']:.0f}% ret {base_ps['ret']:+.0f}%")
print(f" flataware (no-skip) : trd {diag['trades']:>5d} Sh {diag['sharpe']:.2f} "
f"DD {diag['dd']:.0f}% ret {diag['ret']:+.0f}%")
filt = pairs_sim_flataware("ETH", "BTC", **GAME_15M, skip_flat=True)
filt_o = pairs_sim_flataware("ETH", "BTC", **GAME_15M, skip_flat=True, split_frac=1 - OOS_FRAC)
print(f" FLAT-AWARE (skip) : trd {filt['trades']:>5d} Sh {filt['sharpe']:.2f} "
f"DD {filt['dd']:.0f}% ret {filt['ret']:+.0f}% | OOS Sh {filt_o['sharpe']:.2f} DD {filt_o['dd']:.0f}%")
drop = (1 - filt['trades'] / diag['trades']) * 100
sh_keep = filt['sharpe'] / diag['sharpe'] * 100 if diag['sharpe'] else 0
verdict = "EDGE NON artefatto flat" if sh_keep > 70 else "EDGE in larga parte ARTEFATTO flat"
print(f" -> rimossi {drop:.1f}% dei trade; Sharpe trattenuto {sh_keep:.0f}% ({verdict})")
# [4] gate PORT06 col 15m flat-filtrato
print("\n[4] GATE PORT06 — ETH/BTC: baseline 1h vs SWAP 15m-FLATAWARE vs BLEND:")
p = PORTFOLIOS["PORT06"]
pair_ids = [s.sid for s in p.sleeves if s.sid.startswith("PR_")]
eq_base = dict(all_sleeve_equities())
e1h, _ = eth_btc_daily(UNIV_1H)
e15f = daily_from(filt["eq_ts"], filt["eq_v"])
# blend 1h + 15m-flataware (50/50 daily-rebalanced)
from scripts.analysis.pairs15m_port06_gate import blend
eblend = blend(e1h, e15f, 0.5)
corr = e1h.pct_change().fillna(0).corr(e15f.pct_change().fillna(0))
print(f" corr 1h vs 15m-flataware: {corr:.3f}")
print(f" {'variante':<18s} | {'FULL Sh':>8s}{'FULL DD%':>9s}{'CAGR':>6s} | {'OOS Sh':>7s}{'OOS DD%':>8s}")
print(" " + "-" * 70)
res = {}
for tag, eth in [("baseline 1h", e1h), ("SWAP 15m-flat", e15f), ("BLEND 1h+15m-flat", eblend)]:
members = dict(eq_base); members["PR_ETHBTC"] = eth
f, o = port_metrics(members, p)
res[tag] = (f, o)
print(f" {tag:<18s} | {f['sharpe']:>8.2f}{f['dd']:>9.2f}{f['cagr']:>5.0f}%"
f" | {o['sharpe']:>7.2f}{o['dd']:>8.2f}")
fb, ob = res["baseline 1h"]
print("\n VERDETTO (vs baseline 1h, fee backtest): Sharpe non peggiora E DD <= baseline")
for tag in ("SWAP 15m-flat", "BLEND 1h+15m-flat"):
f, o = res[tag]
ok = o["sharpe"] >= ob["sharpe"] - 0.02 and o["dd"] <= ob["dd"] + 1e-9 and f["sharpe"] >= fb["sharpe"] - 0.02 and f["dd"] <= fb["dd"] + 1e-9
print(f" {tag:<18s}: OOS {ob['sharpe']:.2f}->{o['sharpe']:.2f} DD {ob['dd']:.2f}->{o['dd']:.2f}"
f" | FULL {fb['sharpe']:.2f}->{f['sharpe']:.2f} DD {fb['dd']:.2f}->{f['dd']:.2f} => {'PROMOSSO' if ok else 'bocciato'}")
if __name__ == "__main__":
main()
+62
View File
@@ -0,0 +1,62 @@
"""GATE PORT06 FINALE — ETH/BTC 15m flat-skip, engine canonico pairs_sim_flat.
Usa pairs_sim_flat(flat_skip=True), cioe' la STESSA semantica live-realizable del
PairsWorker (uscita alla prima barra pulita), validata da validate_worker_pairs.
Conferma i numeri deployabili: baseline 1h vs SWAP 15m vs BLEND 1h+15m.
uv run python scripts/analysis/pairs15m_gate_final.py
"""
from __future__ import annotations
import sys
from pathlib import Path
PROJECT_ROOT = Path(__file__).resolve().parents[2]
sys.path.insert(0, str(PROJECT_ROOT))
from scripts.analysis.pairs_research import pairs_sim_flat
from scripts.analysis.report_families import daily_from
from scripts.analysis.pairs15m_port06_gate import (port_metrics, eth_btc_daily, blend,
UNIV_1H, POS, LEV)
from scripts.portfolios._defs import PORTFOLIOS
from src.portfolio.sleeves import all_sleeve_equities
CFG_15M = dict(n=66, z_in=1.674, z_exit=1.0, max_bars=35)
def main():
p = PORTFOLIOS["PORT06"]
eq_base = dict(all_sleeve_equities())
e1h, _ = eth_btc_daily(UNIV_1H)
r15 = pairs_sim_flat("ETH", "BTC", tf="15m", **CFG_15M, flat_skip=True, pos=POS, lev=LEV)
e15 = daily_from(r15["eq_ts"], r15["eq_v"])
eblend = blend(e1h, e15, 0.5)
corr = e1h.pct_change().fillna(0).corr(e15.pct_change().fillna(0))
print("=" * 92)
print(" GATE PORT06 FINALE — ETH/BTC 15m flat-skip (pairs_sim_flat == worker live)")
print(f" 15m: {r15['trades']} trade, {r15['n_skip_entry']} ingressi flat saltati | "
f"corr 1h vs 15m = {corr:.3f}")
print("=" * 92)
print(f" {'variante':<18s} | {'FULL Sh':>8s}{'FULL DD%':>9s}{'CAGR':>6s} | {'OOS Sh':>7s}{'OOS DD%':>8s}")
print(" " + "-" * 70)
res = {}
for tag, eth in [("baseline 1h", e1h), ("SWAP 15m-flat", e15), ("BLEND 1h+15m", eblend)]:
members = dict(eq_base); members["PR_ETHBTC"] = eth
f, o = port_metrics(members, p)
res[tag] = (f, o)
print(f" {tag:<18s} | {f['sharpe']:>8.2f}{f['dd']:>9.2f}{f['cagr']:>5.0f}%"
f" | {o['sharpe']:>7.2f}{o['dd']:>8.2f}")
fb, ob = res["baseline 1h"]
print("\n Promosso se OOS Sharpe non peggiora E DD<=baseline (PORT06):")
for tag in ("SWAP 15m-flat", "BLEND 1h+15m"):
f, o = res[tag]
ok = o["sharpe"] >= ob["sharpe"] - 0.02 and o["dd"] <= ob["dd"] + 1e-9 \
and f["sharpe"] >= fb["sharpe"] - 0.02 and f["dd"] <= fb["dd"] + 1e-9
print(f" {tag:<18s}: OOS {ob['sharpe']:.2f}->{o['sharpe']:.2f} DD {ob['dd']:.2f}->{o['dd']:.2f}"
f" | FULL {fb['sharpe']:.2f}->{f['sharpe']:.2f} DD {fb['dd']:.2f}->{f['dd']:.2f}"
f" => {'PROMOSSO' if ok else 'bocciato'}")
if __name__ == "__main__":
main()
+81
View File
@@ -0,0 +1,81 @@
"""Smoke LIVE del nuovo percorso 15m: fetch DIRETTO 15m da Cerbero per ETH/BTC +
freschezza + flat-fraction + un tick reale del PairsWorker(flat_skip).
Verifica cio' che il backtest non vede: che Cerbero serva candele 15m fresche per
entrambe le gambe (il runner ora le fetcha dirette, non resamplate dal 1h) e che il
worker 15m le processi senza errori. NON apre ordini reali (l'esecuzione a 2 gambe e'
gia' coperta da live_pairs_smoke.py, indipendente dal timeframe).
uv run python scripts/analysis/pairs15m_live_smoke.py
"""
from __future__ import annotations
import sys
import shutil
import tempfile
from datetime import datetime, timezone, timedelta
from pathlib import Path
import pandas as pd
PROJECT_ROOT = Path(__file__).resolve().parents[2]
sys.path.insert(0, str(PROJECT_ROOT))
from src.live.cerbero_client import CerberoClient
from src.live.multi_runner import INSTRUMENT_MAP
from src.live.pairs_worker import PairsWorker
CFG = {"n": 66, "z_in": 1.674, "z_exit": 1.0, "max_bars": 35, "flat_skip": True}
def fetch15(cli, asset, days=14):
inst = INSTRUMENT_MAP.get(asset, f"{asset}-PERPETUAL")
end = datetime.now(timezone.utc)
start = end - timedelta(days=days)
candles = cli.get_historical_v2(inst, start.strftime("%Y-%m-%d"),
end.strftime("%Y-%m-%d"), "15m")
if not candles:
return inst, None
df = pd.DataFrame(candles)
df["timestamp"] = df["timestamp"].astype("int64")
return inst, df.sort_values("timestamp").reset_index(drop=True)
def main():
print("=" * 84)
print(" SMOKE LIVE — ETH/BTC pairs 15m (fetch diretto Cerbero + tick worker flat-skip)")
print("=" * 84)
cli = CerberoClient()
inst_a, da = fetch15(cli, "ETH")
inst_b, db = fetch15(cli, "BTC")
ok = True
for asset, inst, df in [("ETH", inst_a, da), ("BTC", inst_b, db)]:
if df is None or df.empty:
print(f" {asset} ({inst}): NESSUNA candela 15m -> FAIL"); ok = False; continue
last = pd.to_datetime(df["timestamp"].iloc[-1], unit="ms", utc=True)
age_min = (datetime.now(timezone.utc) - last).total_seconds() / 60
flat = ((df["open"] == df["high"]) & (df["high"] == df["low"]) &
(df["low"] == df["close"])).mean() * 100
fresh = age_min < 60
print(f" {asset} ({inst}): {len(df)} barre 15m | ultima {last:%Y-%m-%d %H:%M} "
f"({age_min:.0f} min fa, {'FRESCO' if fresh else 'STALE'}) | flat {flat:.1f}%")
ok &= fresh
if da is None or db is None:
print("\n ESITO: FAIL (feed 15m assente)."); return
# tick reale del worker 15m
tmp = Path(tempfile.mkdtemp(prefix="smoke15m_"))
try:
w = PairsWorker("ETH", "BTC", "15m", params=CFG, fee_rt=0.001, data_dir=tmp)
df_a = pd.DataFrame({"timestamp": da["timestamp"], "open": da["open"], "high": da["high"],
"low": da["low"], "close": da["close"]})
df_b = pd.DataFrame({"timestamp": db["timestamp"], "open": db["open"], "high": db["high"],
"low": db["low"], "close": db["close"]})
w.tick(df_a, df_b)
print(f"\n Worker 15m flat_skip={w.flat_skip} -> tick OK | {w.status_summary}")
print(f" ESITO: {'OK — feed 15m fresco e worker ticca' if ok else 'ATTENZIONE: feed 15m stale/parziale'}")
finally:
shutil.rmtree(tmp, ignore_errors=True)
if __name__ == "__main__":
main()
+169
View File
@@ -0,0 +1,169 @@
"""GATE PORT06 — ETH/BTC pairs a 15m (origine: gioco "Blind Traders", vincitore #43).
Domanda onesta sollevata dal gioco: la coppia ETH/BTC (gia' deployata in PR01 a 1h,
config UNIV n=50 z_in=2.0 z_exit=0.75 max_bars=72) MIGLIORA se girata a 15m con la
config trovata dal gioco (n=66 z_in=1.67 z_exit=1.0 max_bars=35), oppure e' solo una
variante piu' veloce, correlata, dello STESSO spread?
Metodo (engine di PRODUZIONE pairs_sim, NON il motore-giocattolo del gioco):
[1] PARITA': pairs_sim ETH/BTC 1h UNIV (pos0.15 lev3) == sleeve canonico PR_ETHBTC.
[2] CORRELAZIONE 1h vs 15m (rendimenti giornalieri): se ~1 e' ridondante.
[3] STANDALONE 1h vs 15m (+ griglia robustezza n x z_in su 15m, + stress fee 2x).
[4] GATE PORT06: baseline(1h) vs SWAP(15m) vs BLEND(0.5*1h+0.5*15m) per la sleeve
ETH/BTC; promosso se vs baseline l'OOS Sharpe non peggiora E il DD scende
(PORT06 e famiglia), come gli altri gate del progetto.
uv run python scripts/analysis/pairs15m_port06_gate.py
"""
from __future__ import annotations
import sys
from pathlib import Path
import numpy as np
import pandas as pd
PROJECT_ROOT = Path(__file__).resolve().parents[2]
sys.path.insert(0, str(PROJECT_ROOT))
from scripts.analysis.combine_portfolio import port_returns, metrics, SPLIT, OOS_DATE, IDX
from scripts.analysis.pairs_research import pairs_sim, OOS_FRAC
from scripts.analysis.report_families import daily_from
from scripts.portfolios._defs import PORTFOLIOS
from src.portfolio import weighting as W
POS, LEV = 0.15, 3.0 # config CANONICA (== build_everything)
UNIV_1H = dict(tf="1h", n=50, z_in=2.0, z_exit=0.75, max_bars=72)
GAME_15M = dict(tf="15m", n=66, z_in=1.674, z_exit=1.0, max_bars=35) # vincitore gioco
def eth_btc_daily(cfg):
r = pairs_sim("ETH", "BTC", **{**cfg, "pos": POS, "lev": LEV})
return daily_from(r["eq_ts"], r["eq_v"]), r
def std_metrics(cfg, fee_rt=0.001):
f = pairs_sim("ETH", "BTC", **{**cfg, "pos": POS, "lev": LEV, "fee_rt": fee_rt})
o = pairs_sim("ETH", "BTC", **{**cfg, "pos": POS, "lev": LEV, "fee_rt": fee_rt,
"split_frac": 1 - OOS_FRAC})
yrs = f["yearly"]; pos_y = sum(1 for v in yrs.values() if v > 0)
return f, o, pos_y, len(yrs)
def port_metrics(members, p):
ids = p.sleeve_ids
dr = pd.DataFrame({i: members[i].pct_change().fillna(0.0) for i in ids})
w = W.weight_vector(p.weighting, ids, dr, weights=p.weights,
caps=p.caps, clusters=p.clusters, lookback=p.vol_lookback)
drp = port_returns({i: members[i] for i in ids}, w)
return metrics(drp), metrics(drp, lo=SPLIT)
def fam_metrics(eqs):
dr = port_returns(eqs)
return metrics(dr), metrics(dr, lo=SPLIT)
def blend(e1, e2, w1=0.5):
"""Sleeve combinata: media pesata dei rendimenti giornalieri (ribilancio 1D)."""
r1 = e1.reindex(IDX).ffill().bfill().pct_change().fillna(0.0)
r2 = e2.reindex(IDX).ffill().bfill().pct_change().fillna(0.0)
rb = w1 * r1 + (1 - w1) * r2
eq = (1 + rb).cumprod()
return eq / eq.iloc[0]
def main():
p = PORTFOLIOS["PORT06"]
pair_ids = [s.sid for s in p.sleeves if s.sid.startswith("PR_")]
print("=" * 100)
print(" GATE PORT06 — ETH/BTC pairs 15m (vincitore gioco) vs 1h deployato")
print(f" pos={POS} lev={LEV} (canonico) | OOS da {OOS_DATE} | coppie PORT06: {pair_ids}")
print("=" * 100)
from src.portfolio.sleeves import all_sleeve_equities
eq_base = dict(all_sleeve_equities())
# [1] PARITA'
print("\n[1] PARITA' pairs_sim ETH/BTC 1h UNIV (pos0.15 lev3) == sleeve canonico PR_ETHBTC:")
e1h, r1h = eth_btc_daily(UNIV_1H)
base = eq_base["PR_ETHBTC"]
corr = base.pct_change().fillna(0).corr(e1h.pct_change().fillna(0))
rb = (base.iloc[-1] / base.iloc[0] - 1) * 100
rr = (e1h.iloc[-1] / e1h.iloc[0] - 1) * 100
par_ok = corr > 0.999 and abs(rr - rb) <= max(1.0, abs(rb) * 0.01)
print(f" corr={corr:.5f} ret canon {rb:+.0f}% vs replay {rr:+.0f}% "
f"{'OK' if par_ok else '<-- MISMATCH (STOP)'}")
if not par_ok:
return
# [2] CORRELAZIONE 1h vs 15m
e15, r15 = eth_btc_daily(GAME_15M)
c = e1h.pct_change().fillna(0).corr(e15.pct_change().fillna(0))
print(f"\n[2] CORRELAZIONE rendimenti giornalieri ETH/BTC 1h vs 15m: {c:.3f}")
print(f" {'(quasi-duplicato se >0.8; diversificatore se <0.5)':<60s}")
# [3] STANDALONE 1h vs 15m
print("\n[3] STANDALONE ETH/BTC (netto fee 0.20% RT/coppia, leva 3x):")
print(f" {'cfg':<10s}{'trd':>6s}{'win%':>6s}{'FULL%':>9s}{'OOS%':>9s}{'CAGR%':>7s}"
f"{'DD%':>6s}{'oDD%':>7s}{'Shrp':>6s}{'anni+':>7s}{'fee2x FULL%':>12s}")
for tag, cfg in [("1h UNIV", UNIV_1H), ("15m gioco", GAME_15M)]:
f, o, py, ny = std_metrics(cfg)
f2, _, _, _ = std_metrics(cfg, fee_rt=0.002)
print(f" {tag:<10s}{f['trades']:>6d}{f['win']:>6.1f}{f['ret']:>+9.0f}{o['ret']:>+9.0f}"
f"{f['cagr']:>7.0f}{f['dd']:>6.0f}{o['dd']:>7.0f}{f['sharpe']:>6.2f}"
f"{f'{py}/{ny}':>7s}{f2['ret']:>+12.0f}")
# robustezza: plateau n x z_in su 15m (Sharpe>1?)
print("\n Robustezza 15m (Sharpe full, griglia n x z_in, z_exit=1.0 max_bars=35):")
ns = [40, 50, 66, 80]; zs = [1.5, 1.7, 2.0, 2.5]
cells = 0; tot = 0
hdr = " n\\z_in " + "".join(f"{z:>7.1f}" for z in zs)
print(hdr)
for n in ns:
row = f" {n:>6d} "
for z in zs:
s = pairs_sim("ETH", "BTC", tf="15m", n=n, z_in=z, z_exit=1.0,
max_bars=35, pos=POS, lev=LEV)["sharpe"]
tot += 1; cells += s > 1
row += f"{s:>7.2f}"
print(row)
print(f" -> {cells}/{tot} celle Sharpe>1 (plateau se ~tutte; picco se poche)")
# [4] GATE PORT06
print("\n[4] GATE PORT06 — sleeve ETH/BTC: baseline(1h) vs SWAP(15m) vs BLEND(50/50):")
variants = {
"baseline 1h": e1h,
"SWAP 15m": e15,
"BLEND 1h+15m": blend(e1h, e15, 0.5),
}
print(f" {'variante':<14s} | {'FULL Sh':>8s}{'FULL DD%':>9s}{'CAGR':>6s}"
f" | {'OOS Sh':>7s}{'OOS DD%':>8s} | {'famSh':>6s}{'famDD%':>7s}")
print(" " + "-" * 78)
res = {}
for tag, eth in variants.items():
members = dict(eq_base)
members["PR_ETHBTC"] = eth
f, o = port_metrics(members, p)
fam_eqs = {sid: (eth if sid == "PR_ETHBTC" else eq_base[sid]) for sid in pair_ids}
ff, _ = fam_metrics(fam_eqs)
res[tag] = (f, o, ff)
print(f" {tag:<14s} | {f['sharpe']:>8.2f}{f['dd']:>9.2f}{f['cagr']:>5.0f}%"
f" | {o['sharpe']:>7.2f}{o['dd']:>8.2f} | {ff['sharpe']:>6.2f}{ff['dd']:>7.1f}")
# VERDETTO
fb, ob, _ = res["baseline 1h"]
print("\n" + "=" * 100)
print(" VERDETTO vs baseline 1h: promosso se OOS Sharpe non peggiora E DD scende (PORT06 e famiglia)")
print("=" * 100)
for tag in ("SWAP 15m", "BLEND 1h+15m"):
f, o, ff = res[tag]
ok = (o["sharpe"] >= ob["sharpe"] - 0.02 and o["dd"] <= ob["dd"] + 1e-9
and f["sharpe"] >= fb["sharpe"] - 0.02)
print(f" {tag:<14s}: OOS Sh {ob['sharpe']:.2f}->{o['sharpe']:.2f} "
f"DD {ob['dd']:.2f}->{o['dd']:.2f} | FULL Sh {fb['sharpe']:.2f}->{f['sharpe']:.2f} "
f"DD {fb['dd']:.2f}->{f['dd']:.2f} => {'PROMOSSO' if ok else 'bocciato'}")
if __name__ == "__main__":
main()
+92
View File
@@ -95,6 +95,98 @@ def pairs_sim(a, b, tf="1h", n=50, z_in=2.0, z_exit=0.5, max_bars=72,
eq_ts=eq_ts, eq_v=eq_v) eq_ts=eq_ts, eq_v=eq_v)
def aligned_ohlc(a: str, b: str, tf: str = "1h"):
"""Come aligned ma con OHLC di ENTRAMBE le gambe (serve a rilevare candele flat)."""
da = load_data(a, tf)[["timestamp", "open", "high", "low", "close"]].rename(
columns=lambda x: x + "_a" if x != "timestamp" else x)
db = load_data(b, tf)[["timestamp", "open", "high", "low", "close"]].rename(
columns=lambda x: x + "_b" if x != "timestamp" else x)
m = da.merge(db, on="timestamp", how="inner").reset_index(drop=True)
m["dt"] = pd.to_datetime(m["timestamp"], unit="ms", utc=True)
return m
def is_flat_ohlc(o, h, l, c):
"""Candela flat (O=H=L=C): prezzo fermo / liquidita' zero -> fill non eseguibile."""
return (o == h) & (h == l) & (l == c)
def pairs_sim_flat(a, b, tf="1h", n=50, z_in=2.0, z_exit=0.5, max_bars=72,
jump_max=0.08, fee_rt=FEE_RT, lev=LEV, pos=POS, split_frac=0.0,
flat_skip=False, scan_buffer=192):
"""Engine pairs GENERALIZZATO con opzione flat-skip LIVE-REALIZABLE.
Identico a pairs_sim quando flat_skip=False (regression-lock verificato).
Con flat_skip=True:
- ENTRY: saltata se la barra d'ingresso e' flat in UNA delle due gambe (prezzo stale).
- EXIT: la condizione di uscita (|z|<=z_exit O bars>=max_bars) arma 'exit_ready';
si esce al CLOSE della PRIMA barra PULITA successiva (mai a un prezzo passato).
scan_buffer = barre extra oltre max_bars concesse per trovare la barra pulita.
Questa e' la stessa regola implementata nel PairsWorker live (flat_skip) -> parita'.
"""
m = aligned_ohlc(a, b, tf)
ca, cb = m["close_a"].values, m["close_b"].values
N = len(ca)
if flat_skip:
flat = (is_flat_ohlc(m["open_a"].values, m["high_a"].values, m["low_a"].values, ca)
| is_flat_ohlc(m["open_b"].values, m["high_b"].values, m["low_b"].values, cb))
else:
flat = np.zeros(N, dtype=bool)
r = np.log(ca / cb)
dr = np.abs(np.diff(r, prepend=r[0]))
ma = pd.Series(r).rolling(n).mean().values
sd = pd.Series(r).rolling(n).std().values
z = (r - ma) / np.where(sd == 0, np.nan, sd)
ts = m["dt"]
split = int(N * split_frac)
fee = 2 * fee_rt * lev
cap = peak = 1000.0; dd = 0.0; last = -1
trades = wins = 0; rets = []; yearly = {}
eq_ts, eq_v = [], []
n_skip_entry = 0
kmax = max_bars + (scan_buffer if flat_skip else 0)
for i in range(n + 1, N - 1):
if i < split or np.isnan(z[i]) or dr[i] > jump_max or i <= last:
continue
if z[i] <= -z_in:
d = 1
elif z[i] >= z_in:
d = -1
else:
continue
if flat[i]:
n_skip_entry += 1
continue # niente ingresso su barra stale
# uscita live-realizable: arma a |z|<=z_exit o max_bars, esci alla prima barra pulita
exit_ready = False; j = i
for k in range(1, kmax + 1):
jj = i + k
if jj >= N:
j = N - 1; break
if not exit_ready and (abs(z[jj]) <= z_exit or k >= max_bars):
exit_ready = True
if exit_ready and not flat[jj]:
j = jj; break
j = jj
retA = (ca[j] - ca[i]) / ca[i]
retB = (cb[j] - cb[i]) / cb[i]
ret = (retA - retB) * d * lev - fee
cap = max(cap + cap * pos * ret, 10.0)
peak = max(peak, cap); dd = max(dd, (peak - cap) / peak)
trades += 1; wins += ret > 0; rets.append(ret * pos); last = j
eq_ts.append(ts.iloc[j]); eq_v.append(cap)
yearly[ts.iloc[i].year] = yearly.get(ts.iloc[i].year, 0.0) + ret * 100
yrs_span = (ts.iloc[-1] - ts.iloc[max(split, 0)]).days / 365.25 or 1
sharpe = 0.0
if len(rets) > 1 and np.std(rets) > 0:
sharpe = float(np.mean(rets) / np.std(rets) * np.sqrt(trades / yrs_span))
ret_tot = (cap / 1000 - 1) * 100
cagr = ((cap / 1000) ** (1 / yrs_span) - 1) * 100 if cap > 0 else -100
return dict(trades=trades, win=wins / trades * 100 if trades else 0, ret=ret_tot,
cagr=cagr, dd=dd * 100, sharpe=sharpe, yearly=yearly,
eq_ts=eq_ts, eq_v=eq_v, n_skip_entry=n_skip_entry)
def check_no_lookahead(): def check_no_lookahead():
"""Perturba il FUTURO del ratio e verifica che z[i] non cambi (causalita').""" """Perturba il FUTURO del ratio e verifica che z[i] non cambi (causalita')."""
m = aligned("ETH", "BTC") m = aligned("ETH", "BTC")
+11 -1
View File
@@ -28,7 +28,7 @@ from scripts.analysis.combine_portfolio import (
build_all_sleeves, port_returns, metrics, yearly_returns, SPLIT, OOS_DATE, IDX, build_all_sleeves, port_returns, metrics, yearly_returns, SPLIT, OOS_DATE, IDX,
) )
from scripts.analysis.honest_improve2 import _daily_equity, _norm from scripts.analysis.honest_improve2 import _daily_equity, _norm
from scripts.analysis.pairs_research import pairs_sim from scripts.analysis.pairs_research import pairs_sim, pairs_sim_flat
from scripts.analysis.tsmom_research import tsmom_sim from scripts.analysis.tsmom_research import tsmom_sim
from scripts.strategies.PR01_pairs_reversion import PAIRS from scripts.strategies.PR01_pairs_reversion import PAIRS
from scripts.analysis.shape_ml_validate import shape_daily_equity from scripts.analysis.shape_ml_validate import shape_daily_equity
@@ -46,6 +46,16 @@ def build_everything():
for a, b, p in PAIRS: for a, b, p in PAIRS:
r = pairs_sim(a, b, **p) r = pairs_sim(a, b, **p)
pairs[f"PR_{a}{b}"] = daily_from(r["eq_ts"], r["eq_v"]) pairs[f"PR_{a}{b}"] = daily_from(r["eq_ts"], r["eq_v"])
# BLEND ETH/BTC 15m flat-skip (gioco Blind Traders -> gate PORT06, decorrelato 0.37
# dal 1h, edge non-artefatto-flat, worker validato). Engine LIVE-REALIZABLE identico
# al PairsWorker (pairs_sim_flat). Diari 2026-06-09-pairs15m-*.md.
# MEZZA size (pos 0.075 = meta' della canonica 0.15): a peso uguale il 15m, piu'
# volatile, contribuirebbe ~26% del rischio PORT06 (vs ~9% del 1h). Dimezzarlo lo
# riporta in linea col 1h -> blend-tilt, non scommessa dominante (col caveat slippage).
# Coerente col live (params.position_size=0.10 = meta' del family PAIRS 0.20).
r15 = pairs_sim_flat("ETH", "BTC", tf="15m", n=66, z_in=1.674, z_exit=1.0,
max_bars=35, flat_skip=True, pos=0.075)
pairs["PR_ETHBTC_15M"] = daily_from(r15["eq_ts"], r15["eq_v"])
t = tsmom_sim() t = tsmom_sim()
tsm = {"TSM01": daily_from(t["eq_ts"], t["eq_v"])} tsm = {"TSM01": daily_from(t["eq_ts"], t["eq_v"])}
shape = {f"SH_{a}": _norm(shape_daily_equity(a, IDX)) for a in ("BTC", "ETH")} shape = {f"SH_{a}": _norm(shape_daily_equity(a, IDX)) for a in ("BTC", "ETH")}
+48 -34
View File
@@ -18,56 +18,70 @@ PROJECT_ROOT = Path(__file__).resolve().parents[2]
sys.path.insert(0, str(PROJECT_ROOT)) sys.path.insert(0, str(PROJECT_ROOT))
from src.live.pairs_worker import PairsWorker from src.live.pairs_worker import PairsWorker
from scripts.analysis.pairs_research import aligned, pairs_sim from scripts.analysis.pairs_research import aligned, aligned_ohlc, pairs_sim, pairs_sim_flat
from scripts.strategies.PR01_pairs_reversion import PAIRS from scripts.strategies.PR01_pairs_reversion import PAIRS
WINDOW = 60 # finestra trailing minima (>= n+2): z[i] corretto, replay veloce # Config 15m promossa dal gate (gioco Blind Traders + flat-skip): vedi
# docs/diary/2026-06-09-pairs15m-port06-gate.md
CFG_15M = dict(n=66, z_in=1.674, z_exit=1.0, max_bars=35, flat_skip=True)
def replay(a: str, b: str, params: dict, data_dir: Path) -> PairsWorker: def replay(a, b, params, data_dir, tf="1h", ohlc=False) -> PairsWorker:
m = aligned(a, b) if ohlc:
df_a = m[["timestamp"]].copy(); df_a["close"] = m["close_a"].values m = aligned_ohlc(a, b, tf)
df_b = m[["timestamp"]].copy(); df_b["close"] = m["close_b"].values df_a = pd.DataFrame({"timestamp": m["timestamp"], "open": m["open_a"],
w = PairsWorker(a, b, "1h", params=params, fee_rt=0.001, data_dir=data_dir) "high": m["high_a"], "low": m["low_a"], "close": m["close_a"]})
# replay veloce: niente I/O su file / log / notifiche ad ogni tick (servono solo le metriche finali) df_b = pd.DataFrame({"timestamp": m["timestamp"], "open": m["open_b"],
"high": m["high_b"], "low": m["low_b"], "close": m["close_b"]})
else:
m = aligned(a, b, tf)
df_a = m[["timestamp"]].copy(); df_a["close"] = m["close_a"].values
df_b = m[["timestamp"]].copy(); df_b["close"] = m["close_b"].values
w = PairsWorker(a, b, tf, params=params, fee_rt=0.001, data_dir=data_dir)
w._save_state = lambda: None w._save_state = lambda: None
w._log = lambda *a, **k: None w._log = lambda *a, **k: None
w._notify = lambda *a, **k: None w._notify = lambda *a, **k: None
n = w.n window = max(60, w.n + 6) # finestra trailing >= n+? : z[i] corretto
for k in range(n + 2, len(m) + 1): for k in range(w.n + 2, len(m) + 1):
lo = max(0, k - WINDOW) lo = max(0, k - window)
w.tick(df_a.iloc[lo:k], df_b.iloc[lo:k]) w.tick(df_a.iloc[lo:k], df_b.iloc[lo:k])
# chiudi eventuale posizione aperta a fine serie (come fa il backtest col troncamento)
return w return w
def _row(label, w, bt):
bt_cap = 1000.0 * (1 + bt["ret"] / 100)
cap_match = abs(w.capital - bt_cap) / bt_cap < 0.02 if bt_cap else False
trd_match = abs(w.total_trades - bt["trades"]) <= max(2, bt["trades"] * 0.02)
ok = "OK" if (cap_match and trd_match) else "DIFF"
ww = w.total_wins / w.total_trades * 100 if w.total_trades else 0
print(f" {label:<16s}{w.capital:>13.0f}{w.total_trades:>6d}{ww:>6.1f} | "
f"{bt_cap:>14.0f}{bt['trades']:>6d}{bt['win']:>6.1f} {ok}")
return ok == "OK"
def main(): def main():
print("=" * 96) print("=" * 100)
print(" VALIDAZIONE PairsWorker — replay live vs backtest pairs_sim (fee 0.20% RT/coppia)") print(" VALIDAZIONE PairsWorker — replay live == backtest (fee 0.20% RT/coppia)")
print("=" * 96) print("=" * 100)
print(f" {'coppia':<10s}{'WORKER cap':>12s}{'trd':>5s}{'win%':>6s} | {'BACKTEST cap':>13s}{'trd':>5s}{'win%':>6s} match?") print(f" {'caso':<16s}{'WORKER cap':>13s}{'trd':>6s}{'win%':>6s} | "
print(" " + "-" * 88) f"{'BACKTEST cap':>14s}{'trd':>6s}{'win%':>6s} match?")
# Sottoinsieme rappresentativo: il codice del worker e' identico per ogni coppia, print(" " + "-" * 92)
# quindi 2 coppie con strutture diverse (alt/major e major/alt) bastano a provare
# l'equivalenza. ~135s/coppia su 73k barre orarie. Per validarle tutte: usa PAIRS.
subset = [pp for pp in PAIRS if (pp[0], pp[1]) in {("ETH", "BTC"), ("BTC", "LTC")}]
tmp = Path(tempfile.mkdtemp(prefix="pairs_validate_")) tmp = Path(tempfile.mkdtemp(prefix="pairs_validate_"))
allok = True
try: try:
for a, b, p in subset: # [A] REGRESSIONE 1h (flat_skip=False, close-only) vs pairs_sim
w = replay(a, b, p, tmp) for a, b, p in [pp for pp in PAIRS if (pp[0], pp[1]) in {("ETH", "BTC"), ("BTC", "LTC")}]:
bt = pairs_sim(a, b, **p) w = replay(a, b, p, tmp, tf="1h", ohlc=False)
bt_cap = 1000.0 * (1 + bt["ret"] / 100) allok &= _row(f"{a}/{b} 1h", w, pairs_sim(a, b, **p))
cap_match = abs(w.capital - bt_cap) / bt_cap < 0.02 if bt_cap else False # [B] NUOVO: 15m flat-skip (OHLC) vs pairs_sim_flat
trd_match = abs(w.total_trades - bt["trades"]) <= max(2, bt["trades"] * 0.02) w = replay("ETH", "BTC", CFG_15M, tmp, tf="15m", ohlc=True)
ok = "OK" if (cap_match and trd_match) else "DIFF" bt = pairs_sim_flat("ETH", "BTC", tf="15m", **CFG_15M)
ww = w.total_wins / w.total_trades * 100 if w.total_trades else 0 allok &= _row("ETH/BTC 15m-flat", w, bt)
print(f" {a+'/'+b:<10s}{w.capital:>12.0f}{w.total_trades:>5d}{ww:>6.1f} | "
f"{bt_cap:>13.0f}{bt['trades']:>5d}{bt['win']:>6.1f} {ok}")
finally: finally:
shutil.rmtree(tmp, ignore_errors=True) shutil.rmtree(tmp, ignore_errors=True)
print(" " + "-" * 88) print(" " + "-" * 92)
print(" match = capitale entro 2% e trade entro 2% del backtest. Differenze minime sono") print(" match = capitale e trade entro 2% del backtest (diff minime = bar finale aperta).")
print(" attese (gestione bar finale/troncamento), ma la semantica deve coincidere.") print(f" ESITO COMPLESSIVO: {'TUTTO OK' if allok else 'DIFFERENZE -> INDAGARE'}")
if __name__ == "__main__": if __name__ == "__main__":
+99
View File
@@ -0,0 +1,99 @@
"""
agent_brief — genera il "digest" ANONIMO che ogni agente cieco riceve.
L'agente non sa che sono BTC/ETH ne' che e' crypto: vede solo due serie X e Y
(rinominate dal motore A/B), una finestra normalizzata (base 100) e statistiche
aggregate. Da queste deve proporre una regola che "anticipi" i movimenti.
Genera anche il MENU dei blocchi (famiglie + range parametri) che l'agente puo'
comporre, in modo che l'output sia una spec backtestabile.
"""
from __future__ import annotations
import json
import numpy as np
from scripts.games.engine import load_anon
def _stats(close, high, low):
r = np.diff(np.log(close))
r = r[np.isfinite(r)]
out = {
"n_bars": int(len(close)),
"ret_vol_pct": round(float(np.std(r) * 100), 4),
"ret_autocorr_lag1": round(float(np.corrcoef(r[:-1], r[1:])[0, 1]), 4),
"ret_autocorr_lag5": round(float(np.corrcoef(r[:-5], r[5:])[0, 1]), 4),
"pct_up_bars": round(float(np.mean(r > 0) * 100), 2),
"skew": round(float(((r - r.mean()) ** 3).mean() / (r.std() ** 3 + 1e-12)), 3),
"kurtosis": round(float(((r - r.mean()) ** 4).mean() / (r.std() ** 4 + 1e-12)), 2),
}
# tendenza a rientrare dopo grandi mosse (|z|>2): segno del rendimento successivo
z = (r - r.mean()) / (r.std() + 1e-12)
big = np.where(np.abs(z[:-1]) > 2)[0]
if len(big) > 20:
nxt = r[big + 1]
same = np.sign(r[big]) == np.sign(nxt)
out["after_big_move_continues_pct"] = round(float(np.mean(same) * 100), 1)
return out
def make_digest(tf: str, window: int = 60, seed: int = 0):
data = load_anon(tf)
n = data["n"]
# finestra recente normalizzata (base 100) per "vedere" la forma
s = max(0, n - window)
dig = {"timeframe_id": {"1h": "T1", "15m": "T2", "5m": "T3"}.get(tf, "T?"),
"n_bars_total": n, "series": {}}
for name in ("A", "B"):
o = data[name]
c = o["close"]
norm = (c[s:] / c[s] * 100.0)
dig["series"][{"A": "X", "B": "Y"}[name]] = {
"stats": _stats(c, o["high"], o["low"]),
"recent_window_norm": [round(float(v), 2) for v in norm],
}
# relazione fra le due serie
ra = np.diff(np.log(data["A"]["close"]))
rb = np.diff(np.log(data["B"]["close"]))
m = min(len(ra), len(rb))
dig["XY_return_correlation"] = round(float(np.corrcoef(ra[:m], rb[:m])[0, 1]), 4)
lr = np.log(data["A"]["close"][:m + 1] / data["B"]["close"][:m + 1])
dig["XY_logratio_ret_autocorr"] = round(
float(np.corrcoef(np.diff(lr)[:-1], np.diff(lr)[1:])[0, 1]), 4)
return dig
MENU = {
"obiettivo": ("Proponi UNA regola che anticipi i movimenti futuri per un PnL "
"netto positivo dopo costi (0.10% andata+ritorno per trade). "
"Servono >=10 operazioni al mese. Non sai cosa siano X e Y."),
"famiglie": {
"zscore": "fade/segui lo z-score del prezzo su 'lookback' barre (entry_thr in sigma)",
"breakout": "rottura del canale max/min su 'lookback' barre (reversion=fade la rottura)",
"ma_cross": "incrocio EMA veloce(lookback)/lenta(lookback*slow_mult)",
"rsi": "RSI(lookback); entry_thr scala le bande attorno a 50",
"momentum": "rendimento su 'lookback' barre vs soglia entry_thr (%)",
"pairs": "market-neutral sullo z del log-rapporto X/Y (long una/short l'altra)",
},
"direzione": ["reversion (vai contro la mossa)", "trend (segui la mossa)"],
"serie": ["X", "Y (solo per single-family)", "pairs usa entrambe"],
"exit": "tp_atr / sl_atr (in unita' ATR), max_bars (durata massima)",
"range": {
"lookback": "5-120", "entry_thr": "1.0-3.5", "tp_atr": "0.5-4.0",
"sl_atr": "1.0-5.0", "max_bars": "6-120", "slow_mult": "2-6",
"exit_thr (pairs)": "0.2-1.0",
},
"output_schema": {
"family": "una di [zscore,breakout,ma_cross,rsi,momentum,pairs]",
"series": "X|Y|AB(pairs)", "direction": "reversion|trend",
"params": "dict coi parametri scelti", "hypothesis": "1-2 frasi: cosa hai notato",
},
}
if __name__ == "__main__":
import sys
tf = sys.argv[1] if len(sys.argv) > 1 else "1h"
print(json.dumps(make_digest(tf), indent=2)[:2000])
+231
View File
@@ -0,0 +1,231 @@
"""
Arena — tournament orchestrator per il gioco "Blind Traders".
100 agenti partono da una spec di strategia (creata alla cieca: vedi
agent_brief.py / workflow). L'orchestratore valuta ogni spec con il backtest
deterministico (engine.evaluate) su TRAIN, da' epoche di elaborazione (ogni
agente affina la propria strategia via hill-climb sui parametri) e OGNI 10
EPOCHE blocca il 10% meno profittevole. Restano i 10 piu' profittevoli.
Punteggio = fitness su PNL + %win, con vincolo >=10 trade/mese (engine).
"""
from __future__ import annotations
import json
import random
from pathlib import Path
import numpy as np
from scripts.games.engine import load_anon, splits3, evaluate
OUT = Path("data/games")
OUT.mkdir(parents=True, exist_ok=True)
# Spazio parametri per famiglia (min, max, tipo)
SPACE = {
"zscore": dict(lookback=(10, 100, "i"), entry_thr=(1.0, 3.5, "f"),
tp_atr=(0.5, 4.0, "f"), sl_atr=(1.0, 5.0, "f"),
max_bars=(6, 72, "i")),
"breakout": dict(lookback=(12, 120, "i"), entry_thr=(0.0, 0.0, "f"),
tp_atr=(0.5, 4.0, "f"), sl_atr=(1.0, 5.0, "f"),
max_bars=(6, 72, "i")),
"ma_cross": dict(lookback=(5, 50, "i"), slow_mult=(2.0, 6.0, "f"),
entry_thr=(0.0, 0.0, "f"), tp_atr=(0.5, 4.0, "f"),
sl_atr=(1.0, 5.0, "f"), max_bars=(6, 72, "i")),
"rsi": dict(lookback=(7, 30, "i"), entry_thr=(1.0, 4.0, "f"),
tp_atr=(0.5, 4.0, "f"), sl_atr=(1.0, 5.0, "f"),
max_bars=(6, 72, "i")),
"momentum": dict(lookback=(6, 72, "i"), entry_thr=(1.0, 6.0, "f"),
tp_atr=(0.5, 4.0, "f"), sl_atr=(1.0, 5.0, "f"),
max_bars=(6, 72, "i")),
"pairs": dict(lookback=(20, 120, "i"), entry_thr=(1.5, 3.0, "f"),
exit_thr=(0.2, 1.0, "f"), max_bars=(24, 120, "i")),
}
SINGLE_FAMILIES = ["zscore", "breakout", "ma_cross", "rsi", "momentum"]
DIRECTIONS = ["reversion", "trend"]
TIMEFRAMES = ["1h", "15m", "5m"] # timing diversi su cui competono gli agenti
def _rand_param(rng, lo, hi, typ):
if typ == "i":
return int(rng.randint(int(lo), int(hi)))
return round(rng.uniform(lo, hi), 3)
def random_spec(rng):
if rng.random() < 0.25:
fam = "pairs"
else:
fam = rng.choice(SINGLE_FAMILIES)
params = {}
for k, (lo, hi, typ) in SPACE[fam].items():
params[k] = _rand_param(rng, lo, hi, typ)
spec = {"family": fam, "params": params, "tf": rng.choice(TIMEFRAMES)}
if fam == "pairs":
spec["series"] = "AB"
else:
spec["series"] = rng.choice(["A", "B"])
spec["params"]["direction"] = rng.choice(DIRECTIONS)
return spec
def mutate(spec, rng, strength=0.25):
"""Perturba la spec (hill-climb). Per lo piu' numerica; raramente
cambia direzione/serie. La famiglia resta fissa (identita' dell'agente)."""
s = json.loads(json.dumps(spec))
fam = s["family"]
# perturba 1-2 parametri numerici
keys = [k for k in SPACE[fam] if SPACE[fam][k][0] != SPACE[fam][k][1]]
for k in rng.sample(keys, k=min(len(keys), rng.randint(1, 2))):
lo, hi, typ = SPACE[fam][k]
cur = s["params"][k]
span = (hi - lo) * strength
nv = cur + rng.uniform(-span, span)
nv = max(lo, min(hi, nv))
s["params"][k] = int(round(nv)) if typ == "i" else round(nv, 3)
if fam != "pairs":
if rng.random() < 0.10:
s["params"]["direction"] = rng.choice(DIRECTIONS)
if rng.random() < 0.05:
s["series"] = rng.choice(["A", "B"])
# il timeframe resta l'identita' dell'agente (timing fisso) -> non muta
return s
def _normalize(spec):
"""Completa/ripulisce una spec proposta da un agente (robustezza)."""
fam = spec.get("family")
if fam not in SPACE:
fam = "zscore"
out = {"family": fam, "params": {}}
for k, (lo, hi, typ) in SPACE[fam].items():
v = spec.get("params", {}).get(k, (lo + hi) / 2)
try:
v = float(v)
except Exception:
v = (lo + hi) / 2
v = max(lo, min(hi, v))
out["params"][k] = int(round(v)) if typ == "i" else round(v, 3)
out["tf"] = spec.get("tf") if spec.get("tf") in TIMEFRAMES else "1h"
if fam == "pairs":
out["series"] = "AB"
else:
out["series"] = spec.get("series", "A") if spec.get("series") in ("A", "B") else "A"
d = spec.get("params", {}).get("direction") or spec.get("direction")
out["params"]["direction"] = d if d in DIRECTIONS else "reversion"
return out
class Agent:
def __init__(self, aid, spec, brief=""):
self.id = aid
self.spec = _normalize(spec)
self.brief = brief # cosa "dice" l'agente (ipotesi NL)
self.train_fit = -1e9 # criterio di hill-climb (l'agente ottimizza qui)
self.valid_fit = -1e9 # criterio dell'orchestratore (cull + rank)
self.metrics = {} # metriche TRAIN
self.vmetrics = {} # metriche VALID
self.alive = True
self.culled_epoch = None
@property
def tf(self):
return self.spec.get("tf", "1h")
def score(self, datasets, splits_map):
data = datasets[self.tf]
tr, va, _ = splits_map[self.tf]
self.metrics = evaluate(data, self.spec, tr)
self.vmetrics = evaluate(data, self.spec, va)
self.train_fit = self.metrics["fitness"]
self.valid_fit = self.vmetrics["fitness"]
def run_tournament(specs, briefs=None, seed=7,
epochs=90, cull_every=10, cull_n=10, log=print):
rng = random.Random(seed)
# carica solo i timeframe effettivamente usati dagli agenti
used_tfs = sorted({_normalize(s).get("tf", "1h") for s in specs})
datasets = {tf: load_anon(tf) for tf in used_tfs}
splits_map = {tf: splits3(datasets[tf], 0.60, 0.20) for tf in used_tfs}
briefs = briefs or [""] * len(specs)
agents = [Agent(i, s, briefs[i] if i < len(briefs) else "")
for i, s in enumerate(specs)]
for a in agents:
a.score(datasets, splits_map)
alive = lambda: [a for a in agents if a.alive]
log(f"[epoch 0] {len(alive())} agenti | best VALID fit "
f"{max(a.valid_fit for a in agents):.1f}")
history = []
for ep in range(1, epochs + 1):
# elaborazione: l'agente affina sul TRAIN (cio' che vede); ricalcola VALID
for a in alive():
cand = mutate(a.spec, rng)
data = datasets[a.tf]
tr, va, _ = splits_map[a.tf]
m = evaluate(data, cand, tr)
if m["fitness"] > a.train_fit:
a.spec = _normalize(cand)
a.metrics, a.train_fit = m, m["fitness"]
a.vmetrics = evaluate(data, a.spec, va)
a.valid_fit = a.vmetrics["fitness"]
# cull ogni N epoche: l'ORCHESTRATORE blocca il 10% meno profittevole
# in VALIDATION (generalizzazione, non overfit sul train)
if ep % cull_every == 0:
av = sorted(alive(), key=lambda a: a.valid_fit)
k = cull_n if len(av) - cull_n >= 10 else max(0, len(av) - 10)
for a in av[:k]:
a.alive = False
a.culled_epoch = ep
log(f"[epoch {ep:2d}] cull {k:2d} -> {len(alive()):3d} vivi | "
f"best VALID {max(a.valid_fit for a in alive()):.1f} | "
f"worst-alive {min(a.valid_fit for a in alive()):.1f}")
history.append({"epoch": ep, "alive": len(alive()),
"best_valid": max(a.valid_fit for a in alive())})
survivors = sorted(alive(), key=lambda a: a.valid_fit, reverse=True)
# report finale: TEST = OOS puro mai toccato dall'ottimizzazione
results = []
for rank, a in enumerate(survivors, 1):
data = datasets[a.tf]
_, _, te = splits_map[a.tf]
test = evaluate(data, a.spec, te)
full = evaluate(data, a.spec, None)
results.append({
"rank": rank, "agent": a.id, "spec": a.spec, "brief": a.brief,
"tf": a.tf, "train": a.metrics, "valid": a.vmetrics,
"test": test, "full": full,
})
payload = {"n_agents": len(specs), "epochs": epochs,
"survivors": len(survivors), "results": results,
"history": history,
"reveal": {"A": "BTC", "B": "ETH", "tf": "1h"}}
(OUT / "tournament_result.json").write_text(json.dumps(payload, indent=2))
return payload
def leaderboard(payload, top=10, log=print):
log("\n================ CLASSIFICA FINALE (top %d) ================" % top)
log("VALID = finestra su cui l'orchestratore giudica | TEST = OOS puro (mai ottimizzato)")
log(f"{'#':>2} {'ag':>4} {'tf':>3} {'famiglia':>9} {'ser':>3} {'dir':>9} "
f"{'TEpnl%':>8} {'TEwin':>5} {'TEtpm':>6} {'TEsh':>5} {'VApnl%':>8} {'VAwin':>5}")
for r in payload["results"][:top]:
sp = r["spec"]; te = r["test"]; va = r["valid"]
d = sp["params"].get("direction", "-")
log(f"{r['rank']:>2} {r['agent']:>4} {sp.get('tf','1h'):>3} {sp['family']:>9} "
f"{sp['series']:>3} {d:>9} {te['pnl_pct']:>8.0f} {te['win_rate']*100:>4.0f}% "
f"{te['tpm']:>6.1f} {te['sharpe']:>5.1f} {va['pnl_pct']:>8.0f} "
f"{va['win_rate']*100:>4.0f}%")
if __name__ == "__main__":
import sys
# modalita' test: 100 agenti random
rng = random.Random(42)
specs = [random_spec(rng) for _ in range(100)]
payload = run_tournament(specs, seed=42)
leaderboard(payload)
+323
View File
@@ -0,0 +1,323 @@
"""
Game engine — "Blind Traders" tournament.
100 agenti ricevono due serie anonime (A, B) — in realta' BTC e ETH 1h — e
propongono strategie senza sapere cosa sono. L'orchestratore (questo motore)
valuta ogni strategia con un backtest deterministico, causale e fee-aware, e
assegna un punteggio su %win + PNL con vincolo >=10 trade/mese.
Tutto causale (nessun look-ahead): i segnali alla barra i usano solo dati
fino a close[i]; l'ingresso e' a close[i], le uscite TP/SL/max_bars intrabar
dalle barre successive.
"""
from __future__ import annotations
import numpy as np
import pandas as pd
from src.data.downloader import load_data
FEE_RT = 0.001 # 0.10% round-trip (taker Deribit, baseline progetto)
TF_BPM = {"5m": 12 * 24 * 30, "15m": 4 * 24 * 30, "1h": 24 * 30} # barre/mese per tf
MIN_TRADES_PER_MONTH = 10.0
# Slippage per LATO (oltre alle fee). 0 = come prima. Single-leg paga 2 lati
# (ingresso+uscita), i pairs ne pagano 4 (2 gambe x 2 lati).
_SLIP = 0.0
def set_slippage(slip_per_side: float):
global _SLIP
_SLIP = float(slip_per_side)
# --------------------------------------------------------------------------
# Dati anonimizzati
# --------------------------------------------------------------------------
def load_anon(tf: str = "1h"):
"""Carica BTC->A, ETH->B allineati sull'intersezione temporale.
Ritorna un dict con array OHLC per A e B + datetime. I nomi reali NON
compaiono: gli agenti vedono solo 'A' e 'B'.
"""
btc = load_data("BTC", tf).copy()
eth = load_data("ETH", tf).copy()
for d in (btc, eth):
d["dt"] = pd.to_datetime(d["datetime"])
btc = btc.set_index("dt")
eth = eth.set_index("dt")
idx = btc.index.intersection(eth.index)
btc = btc.loc[idx].sort_index()
eth = eth.loc[idx].sort_index()
out = {"dt": idx.to_numpy()}
for name, d in (("A", btc), ("B", eth)):
out[name] = {
"open": d["open"].to_numpy(float),
"high": d["high"].to_numpy(float),
"low": d["low"].to_numpy(float),
"close": d["close"].to_numpy(float),
"volume": d["volume"].to_numpy(float),
}
out["n"] = len(idx)
out["tf"] = tf
out["bpm"] = TF_BPM[tf]
return out
# --------------------------------------------------------------------------
# Indicatori causali (vettorizzati)
# --------------------------------------------------------------------------
def _roll_mean(x, w):
return pd.Series(x).rolling(w).mean().to_numpy()
def _roll_std(x, w):
return pd.Series(x).rolling(w).std(ddof=0).to_numpy()
def _ema(x, w):
return pd.Series(x).ewm(span=w, adjust=False).mean().to_numpy()
def _atr(high, low, close, w=14):
pc = np.roll(close, 1)
pc[0] = close[0]
tr = np.maximum(high - low, np.maximum(np.abs(high - pc), np.abs(low - pc)))
return pd.Series(tr).rolling(w).mean().to_numpy()
def _rsi(close, w=14):
d = np.diff(close, prepend=close[0])
up = np.where(d > 0, d, 0.0)
dn = np.where(d < 0, -d, 0.0)
ru = pd.Series(up).ewm(alpha=1 / w, adjust=False).mean().to_numpy()
rd = pd.Series(dn).ewm(alpha=1 / w, adjust=False).mean().to_numpy()
rs = ru / (rd + 1e-12)
return 100 - 100 / (1 + rs)
# --------------------------------------------------------------------------
# Famiglie di segnale -> array di posizione desiderata {-1,0,+1} alla barra i
# (causale: usa solo dati fino a close[i]). +1 = long, -1 = short.
# --------------------------------------------------------------------------
def _signal_single(o, family, p):
"""Segnale per una singola serie. Ritorna (pos_target, atr)."""
close = o["close"]
high, low = o["high"], o["low"]
n = len(close)
atr = _atr(high, low, close, 14)
pos = np.zeros(n)
lb = max(2, int(p["lookback"]))
thr = float(p["entry_thr"])
sign = 1 if p.get("direction", "reversion") == "trend" else -1
if family == "zscore":
ma = _roll_mean(close, lb)
sd = _roll_std(close, lb)
z = (close - ma) / (sd + 1e-12)
pos = np.where(z > thr, sign * -1.0, np.where(z < -thr, sign * 1.0, 0.0))
elif family == "breakout":
hh = pd.Series(high).rolling(lb).max().shift(1).to_numpy()
ll = pd.Series(low).rolling(lb).min().shift(1).to_numpy()
up = close > hh
dn = close < ll
# trend: break-up=long ; reversion: break-up=short
pos = np.where(up, sign * 1.0, np.where(dn, sign * -1.0, 0.0))
elif family == "ma_cross":
fast = _ema(close, lb)
slow = _ema(close, max(lb + 2, int(lb * p.get("slow_mult", 3))))
pos = np.where(fast > slow, sign * 1.0, sign * -1.0)
elif family == "rsi":
r = _rsi(close, lb)
hi = 50 + thr * 10
lo = 50 - thr * 10
pos = np.where(r > hi, sign * -1.0, np.where(r < lo, sign * 1.0, 0.0))
elif family == "momentum":
ret = close / np.roll(close, lb) - 1
ret[:lb] = 0
pos = np.where(ret > thr / 100, sign * 1.0,
np.where(ret < -thr / 100, sign * -1.0, 0.0))
else:
raise ValueError(f"unknown family {family}")
pos = np.nan_to_num(pos)
return pos, atr
# --------------------------------------------------------------------------
# Backtest single-series (long/short con TP/SL/max_bars intrabar)
# --------------------------------------------------------------------------
def _backtest_single(o, pos, atr, p, fee=FEE_RT):
close, high, low = o["close"], o["high"], o["low"]
n = len(close)
tp_atr = float(p.get("tp_atr", 2.0))
sl_atr = float(p.get("sl_atr", 2.0))
max_bars = int(p.get("max_bars", 24))
rets = [] # net return per trade
# warmup
start = max(int(p["lookback"]) + 15, 20)
# indici candidati: solo barre con segnale != 0 (salta le barre flat)
cand = np.flatnonzero(pos[start:n - 1]) + start
ci = 0
nc = len(cand)
while ci < nc:
i = int(cand[ci])
d = pos[i]
if d == 0 or np.isnan(atr[i]) or atr[i] <= 0:
ci += 1
continue
entry = close[i]
a = atr[i]
if d > 0:
tp = entry + tp_atr * a
sl = entry - sl_atr * a
else:
tp = entry - tp_atr * a
sl = entry + sl_atr * a
exit_px = None
j = i + 1
end = min(n - 1, i + max_bars)
while j <= end:
hi, lo = high[j], low[j]
if d > 0:
if lo <= sl: # SL prioritario
exit_px = sl
break
if hi >= tp:
exit_px = tp
break
else:
if hi >= sl:
exit_px = sl
break
if lo <= tp:
exit_px = tp
break
j += 1
if exit_px is None:
exit_px = close[end]
j = end
gross = d * (exit_px - entry) / entry
net = gross - fee - 2 * _SLIP # 2 lati di slippage
rets.append(net)
# salta al primo ingresso candidato OLTRE l'uscita (no overlap)
ci = int(np.searchsorted(cand, j + 1, side="left"))
return np.array(rets)
# --------------------------------------------------------------------------
# Backtest cross-series (pairs market-neutral sullo z del log-ratio)
# --------------------------------------------------------------------------
def _backtest_pairs(A, B, p, fee=FEE_RT):
a, b = A["close"], B["close"]
n = len(a)
lb = max(5, int(p["lookback"]))
z_in = float(p["entry_thr"])
z_exit = float(p.get("exit_thr", 0.5))
max_bars = int(p.get("max_bars", 72))
lr = np.log(a / b)
ma = _roll_mean(lr, lb)
sd = _roll_std(lr, lb)
z = (lr - ma) / (sd + 1e-12)
rets = []
start = max(lb + 5, 20)
zabs = np.abs(z)
zabs[:start] = 0.0
zabs[np.isnan(zabs)] = 0.0
cand = np.flatnonzero(zabs[:n - 1] > z_in)
ci = 0
nc = len(cand)
while ci < nc:
i = int(cand[ci])
d = -1 if z[i] > z_in else 1 # spread alto -> short A/long B ; basso -> long A/short B
ea, eb = a[i], b[i]
j = i + 1
end = min(n - 1, i + max_bars)
while j <= end:
if abs(z[j]) <= z_exit:
break
j += 1
j = min(j, end)
# PnL = gamba A (dir d) + gamba B (dir -d), fee su 2 gambe
ra = d * (a[j] - ea) / ea
rb = -d * (b[j] - eb) / eb
net = ra + rb - 2 * fee - 4 * _SLIP # 2 gambe x 2 lati di slippage
rets.append(net)
ci = int(np.searchsorted(cand, j + 1, side="left"))
return np.array(rets)
# --------------------------------------------------------------------------
# Valutazione + scoring
# --------------------------------------------------------------------------
def evaluate(data, spec, sl=None, fee=FEE_RT):
"""Valuta una spec di strategia su uno slice [start,end) (sl=slice di indici).
spec = {family, series, params{...}}. Ritorna dict metriche.
"""
family = spec["family"]
series = spec.get("series", "A")
p = spec["params"]
def _slice(o):
if sl is None:
return o
s, e = sl
return {k: v[s:e] for k, v in o.items()}
if family == "pairs":
A = _slice(data["A"])
B = _slice(data["B"])
rets = _backtest_pairs(A, B, p, fee)
nbars = len(A["close"])
else:
o = _slice(data[series])
pos, atr = _signal_single(o, family, p)
rets = _backtest_single(o, pos, atr, p, fee)
nbars = len(o["close"])
n_tr = len(rets)
months = nbars / data.get("bpm", TF_BPM["1h"])
tpm = n_tr / months if months > 0 else 0.0
if n_tr == 0:
return dict(n_trades=0, win_rate=0.0, pnl_pct=0.0, tpm=0.0,
sharpe=0.0, avg_ret=0.0, qualified=False, fitness=-1e6)
win_rate = float(np.mean(rets > 0))
pnl = float(np.sum(rets)) * 100 # PnL additivo (notional fisso), %
equity = float(np.prod(1 + rets) - 1) * 100 # equity compounding, %
avg = float(np.mean(rets)) * 100
sharpe = float(np.mean(rets) / (np.std(rets) + 1e-12) * np.sqrt(tpm * 12)) \
if np.std(rets) > 0 else 0.0
qualified = tpm >= MIN_TRADES_PER_MONTH
# fitness: PNL domina, win% come spinta secondaria; squalifica se pochi trade
fitness = pnl + 50.0 * win_rate
if not qualified:
fitness = -1e6 + pnl # ordinati ma fuori gioco
return dict(n_trades=n_tr, win_rate=win_rate, pnl_pct=pnl, equity_pct=equity,
tpm=tpm, sharpe=sharpe, avg_ret=avg, qualified=qualified,
fitness=fitness)
# Split a 3: TRAIN (hill-climb) / VALID (cull+rank dell'orchestratore) / TEST (OOS puro)
def splits3(data, train_frac=0.60, valid_frac=0.20):
n = data["n"]
c1 = int(n * train_frac)
c2 = int(n * (train_frac + valid_frac))
return (0, c1), (c1, c2), (c2, n)
# compat: split a 2 (train/oos)
def splits(data, train_frac=0.70):
n = data["n"]
cut = int(n * train_frac)
return (0, cut), (cut, n)
if __name__ == "__main__":
data = load_anon("1h")
print("loaded", data["n"], "bars,", data["dt"][0], "->", data["dt"][-1])
tr, oos = splits(data)
demo = {"family": "zscore", "series": "B",
"params": {"lookback": 20, "entry_thr": 2.0, "direction": "reversion",
"tp_atr": 1.5, "sl_atr": 2.0, "max_bars": 24}}
print("TRAIN", evaluate(data, demo, tr))
print("OOS ", evaluate(data, demo, oos))
+83
View File
@@ -0,0 +1,83 @@
"""
run_game — carica le 100 strategie proposte dagli agenti ciechi (file in
data/games/specs/agent_*.json), lancia il torneo (epoche + cull) e stampa la
classifica finale, poi RIVELA cosa erano X e Y.
Se mancano agenti (file assenti o malformati) riempie con spec casuali, cosi'
il gioco gira sempre a 100 concorrenti.
"""
from __future__ import annotations
import json
import os
import random
from pathlib import Path
from scripts.games import engine
from scripts.games.arena import random_spec, run_tournament, leaderboard, _normalize
SPECS_DIR = Path("data/games/specs")
N = 100
def load_specs():
rng = random.Random(123)
specs, briefs, sources = [], [], []
for i in range(N):
f = SPECS_DIR / f"agent_{i}.json"
spec = None
if f.exists():
try:
raw = json.loads(f.read_text())
fam = raw.get("family")
params = dict(raw.get("params", {}))
if "direction" in raw and "direction" not in params:
params["direction"] = raw["direction"]
spec = {"family": fam, "series": raw.get("series", "A"),
"tf": raw.get("tf", "1h"), "params": params}
# X->A, Y->B mapping (gli agenti vedono X/Y)
s = spec["series"]
spec["series"] = {"X": "A", "Y": "B", "AB": "AB",
"A": "A", "B": "B"}.get(s, "A")
spec = _normalize(spec)
briefs.append(str(raw.get("hypothesis", ""))[:300])
sources.append("agent")
except Exception as e:
spec = None
if spec is None:
spec = random_spec(rng)
briefs.append("(spec mancante -> sostituto casuale)")
sources.append("random")
specs.append(spec)
n_agent = sources.count("agent")
print(f"caricati {n_agent}/{N} spec da agenti reali, "
f"{N - n_agent} sostituiti casuali")
return specs, briefs
def main():
slip = float(os.environ.get("GAME_SLIP", "0.0"))
engine.set_slippage(slip)
if slip > 0:
print(f"SLIPPAGE attivo: {slip*100:.3f}%/lato "
f"(single-leg {2*slip*100:.2f}% RT extra, pairs {4*slip*100:.2f}% extra)")
specs, briefs = load_specs()
payload = run_tournament(specs, briefs=briefs, seed=2026,
epochs=90, cull_every=10, cull_n=10)
leaderboard(payload, top=10)
rev = payload["reveal"]
print(f"\n>>> RIVELAZIONE: Serie X = {rev['A']}, Serie Y = {rev['B']} "
f"(timeframe base {rev['tf']}). Gli agenti non lo sapevano. <<<")
# vincitore
w = payload["results"][0]
sp = w["spec"]
print(f"\nVINCITORE: agente #{w['agent']} su {w['tf']} | {sp['family']} "
f"{sp['series']} {sp['params'].get('direction','')}")
print(f" ipotesi dell'agente: {w['brief']}")
print(f" TEST(OOS): PnL {w['test']['pnl_pct']:.0f}% | win "
f"{w['test']['win_rate']*100:.0f}% | {w['test']['tpm']:.1f} trade/mese "
f"| Sharpe {w['test']['sharpe']:.1f}")
if __name__ == "__main__":
main()
+7
View File
@@ -79,6 +79,13 @@ HONEST = [
] ]
PAIRS = [ PAIRS = [
SleeveSpec(kind="pairs", name="PR01", sid="PR_ETHBTC", a="ETH", b="BTC", cluster="ETH-rev"), SleeveSpec(kind="pairs", name="PR01", sid="PR_ETHBTC", a="ETH", b="BTC", cluster="ETH-rev"),
# BLEND timeframe: ETH/BTC anche a 15m (flat-skip), accanto al 1h. Decorrelato 0.37 dal
# 1h -> diversificatore intra-pairs. Worker validato (validate_worker_pairs 15m, replay
# == pairs_sim_flat). Gate PORT06: docs/diary/2026-06-09-pairs15m-live-path.md.
SleeveSpec(kind="pairs", name="PR01", sid="PR_ETHBTC_15M", a="ETH", b="BTC", tf="15m",
params={"n": 66, "z_in": 1.674, "z_exit": 1.0, "max_bars": 35, "flat_skip": True,
"position_size": 0.10}, # meta' del family PAIRS (0.20): blend-tilt
cluster="ETH-rev"),
SleeveSpec(kind="pairs", name="PR01", sid="PR_LTCETH", a="LTC", b="ETH", cluster="ETH-rev"), SleeveSpec(kind="pairs", name="PR01", sid="PR_LTCETH", a="LTC", b="ETH", cluster="ETH-rev"),
SleeveSpec(kind="pairs", name="PR01", sid="PR_ADAETH", a="ADA", b="ETH", cluster="ETH-rev"), SleeveSpec(kind="pairs", name="PR01", sid="PR_ADAETH", a="ADA", b="ETH", cluster="ETH-rev"),
SleeveSpec(kind="pairs", name="PR01", sid="PR_BTCLTC", a="BTC", b="LTC", cluster="BTC-rev"), SleeveSpec(kind="pairs", name="PR01", sid="PR_BTCLTC", a="BTC", b="LTC", cluster="BTC-rev"),
+34 -11
View File
@@ -51,6 +51,11 @@ class PairsWorker:
self.z_exit = float(p.get("z_exit", 0.75)) self.z_exit = float(p.get("z_exit", 0.75))
self.max_bars = int(p.get("max_bars", 72)) self.max_bars = int(p.get("max_bars", 72))
self.jump_max = float(p.get("jump_max", 0.08)) self.jump_max = float(p.get("jump_max", 0.08))
# flat-skip (timeframe sub-orari, es. 15m): non entrare/uscire su candele flat
# (O=H=L=C, prezzo stale/liquidita' zero -> fill non eseguibile). LIVE-REALIZABLE:
# l'uscita arma exit_ready e si esegue alla prima barra PULITA. Parita' col backtest
# pairs_research.pairs_sim_flat(flat_skip=True). Default off = comportamento 1h storico.
self.flat_skip = bool(p.get("flat_skip", False))
self.initial_capital = capital self.initial_capital = capital
self.position_size = position_size self.position_size = position_size
@@ -71,6 +76,7 @@ class PairsWorker:
self.entry_z = 0.0 self.entry_z = 0.0
self.entry_time = "" self.entry_time = ""
self.bars_held = 0 self.bars_held = 0
self.exit_ready = False # flat-skip: condizione di uscita armata, attende barra pulita
self.total_trades = 0 self.total_trades = 0
self.total_wins = 0 self.total_wins = 0
self.last_bar_ts = 0 self.last_bar_ts = 0
@@ -117,6 +123,7 @@ class PairsWorker:
self.entry_z = s.get("entry_z", 0.0) self.entry_z = s.get("entry_z", 0.0)
self.entry_time = s.get("entry_time", "") self.entry_time = s.get("entry_time", "")
self.bars_held = s.get("bars_held", 0) self.bars_held = s.get("bars_held", 0)
self.exit_ready = s.get("exit_ready", False)
self.total_trades = s.get("total_trades", 0) self.total_trades = s.get("total_trades", 0)
self.total_wins = s.get("total_wins", 0) self.total_wins = s.get("total_wins", 0)
self.last_bar_ts = s.get("last_bar_ts", 0) self.last_bar_ts = s.get("last_bar_ts", 0)
@@ -145,7 +152,8 @@ class PairsWorker:
"capital": round(self.capital, 2), "in_position": self.in_position, "capital": round(self.capital, 2), "in_position": self.in_position,
"direction": self.direction, "entry_a": self.entry_a, "entry_b": self.entry_b, "direction": self.direction, "entry_a": self.entry_a, "entry_b": self.entry_b,
"entry_z": round(self.entry_z, 4), "entry_time": self.entry_time, "entry_z": round(self.entry_z, 4), "entry_time": self.entry_time,
"bars_held": self.bars_held, "total_trades": self.total_trades, "bars_held": self.bars_held, "exit_ready": self.exit_ready,
"total_trades": self.total_trades,
"total_wins": self.total_wins, "last_bar_ts": self.last_bar_ts, "total_wins": self.total_wins, "last_bar_ts": self.last_bar_ts,
"started_at": self.started_at, "last_update": datetime.now(timezone.utc).isoformat(), "started_at": self.started_at, "last_update": datetime.now(timezone.utc).isoformat(),
"real_capital": round(self.real_capital, 4), "real_in_position": self.real_in_position, "real_capital": round(self.real_capital, 4), "real_in_position": self.real_in_position,
@@ -185,6 +193,7 @@ class PairsWorker:
self.entry_a, self.entry_b, self.entry_z = ca, cb, z self.entry_a, self.entry_b, self.entry_z = ca, cb, z
self.entry_time = datetime.now(timezone.utc).isoformat() self.entry_time = datetime.now(timezone.utc).isoformat()
self.bars_held = 0 self.bars_held = 0
self.exit_ready = False
data = {"direction": "long_ratio" if d == 1 else "short_ratio", data = {"direction": "long_ratio" if d == 1 else "short_ratio",
"long_leg": self.asset_a if d == 1 else self.asset_b, "long_leg": self.asset_a if d == 1 else self.asset_b,
"short_leg": self.asset_b if d == 1 else self.asset_a, "short_leg": self.asset_b if d == 1 else self.asset_a,
@@ -287,9 +296,13 @@ class PairsWorker:
"""Chiamato ad ogni poll con gli OHLCV aggiornati delle due gambe.""" """Chiamato ad ogni poll con gli OHLCV aggiornati delle due gambe."""
if df_a is None or df_b is None or df_a.empty or df_b.empty: if df_a is None or df_b is None or df_a.empty or df_b.empty:
return return
m = df_a[["timestamp", "close"]].rename(columns={"close": "ca"}).merge( # merge OHLC quando disponibile (serve a rilevare le candele flat per il flat-skip);
df_b[["timestamp", "close"]].rename(columns={"close": "cb"}), on="timestamp", how="inner" # se le colonne OHLC mancano, flat resta False -> comportamento close-only invariato.
).sort_values("timestamp").reset_index(drop=True) ohlc = ["open", "high", "low", "close"]
keep_a = ["timestamp"] + [c for c in ohlc if c in df_a.columns]
keep_b = ["timestamp"] + [c for c in ohlc if c in df_b.columns]
m = df_a[keep_a].merge(df_b[keep_b], on="timestamp", how="inner",
suffixes=("_a", "_b")).sort_values("timestamp").reset_index(drop=True)
# Scarta la barra IN FORMAZIONE: entry ED exit valutati SOLO sul close di # Scarta la barra IN FORMAZIONE: entry ED exit valutati SOLO sul close di
# barra COMPLETA, come il backtest (pairs_research: close settled) — # barra COMPLETA, come il backtest (pairs_research: close settled) —
# lezione EXIT-16. Detection condivisa: src.live.bars. # lezione EXIT-16. Detection condivisa: src.live.bars.
@@ -298,7 +311,7 @@ class PairsWorker:
m = m.iloc[:-1] m = m.iloc[:-1]
if len(m) < self.n + 2: if len(m) < self.n + 2:
return return
ca, cb = m["ca"].values, m["cb"].values ca, cb = m["close_a"].values, m["close_b"].values
z, dr = self._zscore(ca, cb) z, dr = self._zscore(ca, cb)
i = len(m) - 1 i = len(m) - 1
cur_ts = int(m["timestamp"].iloc[i]) cur_ts = int(m["timestamp"].iloc[i])
@@ -306,19 +319,29 @@ class PairsWorker:
if np.isnan(zi): if np.isnan(zi):
self._save_state(); return self._save_state(); return
# flat della barra corrente (entrambe le gambe): O=H=L=C in una delle due
flat_i = False
if self.flat_skip and {"open_a", "high_a", "low_a"}.issubset(m.columns) \
and {"open_b", "high_b", "low_b"}.issubset(m.columns):
fa = (m["open_a"].iloc[i] == m["high_a"].iloc[i] == m["low_a"].iloc[i] == ca[i])
fb = (m["open_b"].iloc[i] == m["high_b"].iloc[i] == m["low_b"].iloc[i] == cb[i])
flat_i = bool(fa or fb)
if self.in_position: if self.in_position:
if cur_ts > self.last_bar_ts: if cur_ts > self.last_bar_ts:
self.bars_held += 1 self.bars_held += 1
self.last_bar_ts = cur_ts self.last_bar_ts = cur_ts
if abs(zi) <= self.z_exit: # arma l'uscita: |z|<=z_exit (rientro) o time-limit; poi esegui alla 1a barra pulita
self._close(float(ca[i]), float(cb[i]), float(zi), "mean_revert") if not self.exit_ready and (abs(zi) <= self.z_exit or self.bars_held >= self.max_bars):
elif self.bars_held >= self.max_bars: self.exit_ready = True
self._close(float(ca[i]), float(cb[i]), float(zi), "time_limit") if self.exit_ready and not flat_i:
reason = "mean_revert" if abs(zi) <= self.z_exit else "time_limit"
self._close(float(ca[i]), float(cb[i]), float(zi), reason)
self._save_state() self._save_state()
return return
# flat: cerca ingresso (no look-ahead: z[i] usa solo dati <= i) # cerca ingresso (no look-ahead: z[i] usa solo dati <= i); mai su barra stale
if dr[i] <= self.jump_max: if dr[i] <= self.jump_max and not flat_i:
if zi <= -self.z_in: if zi <= -self.z_in:
self._open(1, float(ca[i]), float(cb[i]), float(zi)); self.last_bar_ts = cur_ts self._open(1, float(ca[i]), float(cb[i]), float(zi)); self.last_bar_ts = cur_ts
elif zi >= self.z_in: elif zi >= self.z_in:
+45 -9
View File
@@ -34,16 +34,22 @@ _MULTI_KINDS = ("basket", "rotation", "tsmom")
DATA_DIR = Path("data/portfolio_paper") DATA_DIR = Path("data/portfolio_paper")
# giorni di storia da fetchare per timeframe (TSM01 1d usa 252 barre -> ~440 giorni col buffer) # giorni di storia da fetchare per timeframe (TSM01 1d usa 252 barre -> ~440 giorni col buffer)
_LOOKBACK_DAYS = {"1h": 90, "4h": 220, "1d": 440} _LOOKBACK_DAYS = {"5m": 7, "15m": 14, "30m": 21, "1h": 90, "4h": 220, "1d": 440}
# timeframe SUB-orari: si fetchano DIRETTI da Cerbero (non resamplabili dal 1h).
_SUBHOURLY = {"5m", "15m", "30m"}
# SH01 (ml) richiede >=4000 barre 1h (train_min di ml_wf_entries); 365g (~8760 barre) danno # SH01 (ml) richiede >=4000 barre 1h (train_min di ml_wf_entries); 365g (~8760 barre) danno
# margine ampio per il walk-forward. Difensivo: non dipende dal fetch 440g di TSM01/ROT02. # margine ampio per il walk-forward. Difensivo: non dipende dal fetch 440g di TSM01/ROT02.
_ML_LOOKBACK_DAYS = 365 _ML_LOOKBACK_DAYS = 365
def pos_for_spec(sid: str, global_ps: float, family_overrides: dict[str, float]) -> float: def pos_for_spec(sid: str, global_ps: float, family_overrides: dict[str, float],
"""position_size effettivo di uno sleeve: override per-famiglia (chiave = sleeve_ps: float | None = None) -> float:
weighting.family_of: PAIRS/FADE/HONEST/SHAPE/TSM) o globale.""" """position_size effettivo di uno sleeve. Precedenza: override PER-SLEEVE
(spec.params['position_size'], es. il 15m a 0.10) > override per-FAMIGLIA
(weighting.family_of: PAIRS/FADE/...) > globale."""
from src.portfolio.weighting import family_of from src.portfolio.weighting import family_of
if sleeve_ps is not None:
return float(sleeve_ps)
return family_overrides.get(family_of(sid), global_ps) return family_overrides.get(family_of(sid), global_ps)
@@ -301,7 +307,7 @@ def run(config_path: str = "portfolios.yml"):
ex, inst = _exec_for(s) ex, inst = _exec_for(s)
pex, pinst = _pairs_exec_for(s) pex, pinst = _pairs_exec_for(s)
workers[s.sid] = build_worker_for(s, alloc[s.sid], p.leverage, workers[s.sid] = build_worker_for(s, alloc[s.sid], p.leverage,
position_size=pos_for_spec(s.sid, position_size, ps_family), position_size=pos_for_spec(s.sid, position_size, ps_family, s.params.get("position_size")),
executor=ex, exec_instrument=inst, executor=ex, exec_instrument=inst,
pairs_executor=pex, exec_instruments=pinst) pairs_executor=pex, exec_instruments=pinst)
if ps_family: if ps_family:
@@ -312,7 +318,7 @@ def run(config_path: str = "portfolios.yml"):
paper_dir = DATA_DIR.parent / "portfolio_paper_stats" paper_dir = DATA_DIR.parent / "portfolio_paper_stats"
paper_workers = {s.sid: build_worker_for(s, paper_notional, p.leverage, paper_workers = {s.sid: build_worker_for(s, paper_notional, p.leverage,
data_dir=paper_dir, data_dir=paper_dir,
position_size=pos_for_spec(s.sid, position_size, ps_family)) position_size=pos_for_spec(s.sid, position_size, ps_family, s.params.get("position_size")))
for s in paper_specs} for s in paper_specs}
# bootstrap storia full per gli sleeve ML (SH01): parquet locale + feed live. # bootstrap storia full per gli sleeve ML (SH01): parquet locale + feed live.
@@ -340,6 +346,18 @@ def run(config_path: str = "portfolios.yml"):
for a in assets: for a in assets:
asset_days[a] = max(asset_days.get(a, 0), days) asset_days[a] = max(asset_days.get(a, 0), days)
# timeframe SUB-orari (es. pairs 15m, flat-skip): non resamplabili dal 1h ->
# fetch DIRETTO da Cerbero per (asset, tf). Inerte se nessuno sleeve e' sub-orario.
subhourly_needs: dict[tuple[str, str], int] = {}
for s in supported: # live + paper
assets, tf = _spec_assets_tf(s)
if tf in _SUBHOURLY:
for a in assets:
subhourly_needs[(a, tf)] = max(subhourly_needs.get((a, tf), 0),
_LOOKBACK_DAYS.get(tf, 14))
if subhourly_needs:
print(f"[runner] timeframe sub-orari (fetch diretto Cerbero): {sorted(subhourly_needs)}")
inst_map = dict(INSTRUMENT_MAP) inst_map = dict(INSTRUMENT_MAP)
last_day = "" last_day = ""
stale_alerted: set[str] = set() # asset con alert STALE_FEED attivo (dedup per episodio) stale_alerted: set[str] = set() # asset con alert STALE_FEED attivo (dedup per episodio)
@@ -394,12 +412,30 @@ def run(config_path: str = "portfolios.yml"):
raw1h[asset] = df.sort_values("timestamp").reset_index(drop=True) raw1h[asset] = df.sort_values("timestamp").reset_index(drop=True)
_check_stale_feed(asset, raw1h[asset], stale_alerted) _check_stale_feed(asset, raw1h[asset], stale_alerted)
# tick di ogni worker col suo timeframe (resample dal 1h) # fetch DIRETTO dei timeframe sub-orari (15m...) per (asset, tf)
raw_sub: dict[tuple[str, str], pd.DataFrame] = {}
for (asset, tf), days in subhourly_needs.items():
inst = inst_map.get(asset, f"{asset}-PERPETUAL")
start = end - timedelta(days=days)
candles = client.get_historical_v2(inst, start.strftime("%Y-%m-%d"),
end.strftime("%Y-%m-%d"), tf)
if candles:
df = pd.DataFrame(candles)
df["timestamp"] = df["timestamp"].astype("int64")
raw_sub[(asset, tf)] = df.sort_values("timestamp").reset_index(drop=True)
def _series_for(a, tf):
"""Serie OHLC per (asset, tf): diretta se sub-oraria, altrimenti resample dal 1h."""
if tf in _SUBHOURLY:
return raw_sub.get((a, tf))
return _resample(raw1h[a], tf) if a in raw1h else None
# tick di ogni worker col suo timeframe
def _tick(s, w): def _tick(s, w):
assets, tf = _spec_assets_tf(s) assets, tf = _spec_assets_tf(s)
if any(a not in raw1h for a in assets): res = {a: _series_for(a, tf) for a in assets}
if any(res[a] is None or len(res[a]) == 0 for a in assets):
return return
res = {a: _resample(raw1h[a], tf) for a in assets}
if s.kind == "pairs": if s.kind == "pairs":
w.tick(res[s.a], res[s.b]) w.tick(res[s.a], res[s.b])
elif s.kind in _MULTI_KINDS: elif s.kind in _MULTI_KINDS:
+6 -3
View File
@@ -8,6 +8,9 @@ def test_port06_cap_backtest_numbers_locked():
# Aggiornato 2026-05-31: il recupero dati BNB/DOGE/XRP (29 mag) ha ampliato la # Aggiornato 2026-05-31: il recupero dati BNB/DOGE/XRP (29 mag) ha ampliato la
# copertura storica -> metriche migliorate (Sharpe 6.07->6.47, OOS 8.19->8.82, # copertura storica -> metriche migliorate (Sharpe 6.07->6.47, OOS 8.19->8.82,
# DD 4.9%->4.1%). Nuovo baseline atteso, non una regressione. # DD 4.9%->4.1%). Nuovo baseline atteso, non una regressione.
assert r.full["sharpe"] == pytest.approx(6.47, abs=0.15) # Aggiornato 2026-06-09: aggiunto lo sleeve BLEND PR_ETHBTC_15M (ETH/BTC pairs 15m
assert r.oos["sharpe"] == pytest.approx(8.82, abs=0.25) # flat-skip, mezza size) -> miglioria attesa: FULL 6.47->7.20, OOS 8.82->9.66,
assert r.full["dd"] == pytest.approx(4.1, abs=0.5) # DD 4.1%->3.7%. Vedi docs/diary/2026-06-09-pairs15m-live-path.md.
assert r.full["sharpe"] == pytest.approx(7.20, abs=0.15)
assert r.oos["sharpe"] == pytest.approx(9.66, abs=0.25)
assert r.full["dd"] == pytest.approx(3.68, abs=0.5)
+3 -2
View File
@@ -8,8 +8,9 @@ def test_six_portfolios_defined():
def test_port06_is_master_shape_cap(): def test_port06_is_master_shape_cap():
p = PORTFOLIOS["PORT06"] p = PORTFOLIOS["PORT06"]
sids = set(p.sleeve_ids) sids = set(p.sleeve_ids)
assert {"SH_BTC", "SH_ETH", "TSM01", "PR_ETHBTC"} <= sids assert {"SH_BTC", "SH_ETH", "TSM01", "PR_ETHBTC", "PR_ETHBTC_15M"} <= sids
assert len(sids) == 17 # 18 dal 2026-06-09: aggiunto lo sleeve BLEND PR_ETHBTC_15M (ETH/BTC pairs 15m flat-skip)
assert len(sids) == 18
# SHAPE cappata a 0.0588 (2026-06-05): SH01 senza SL by-design, esposizione dimezzata # SHAPE cappata a 0.0588 (2026-06-05): SH01 senza SL by-design, esposizione dimezzata
# (ricerca sh01_exit_lab: 11 famiglie di stop, 0 sopravvissute) # (ricerca sh01_exit_lab: 11 famiglie di stop, 0 sopravvissute)
assert p.weighting == "cap" and p.caps == {"PAIRS": 0.33, "SHAPE": 0.0588} assert p.weighting == "cap" and p.caps == {"PAIRS": 0.33, "SHAPE": 0.0588}