research(alt): sweep 104 strategie alternative su Deribit (153 agenti) + marginal scorer
Ondata di ricerca onesta a largo spettro su BTC/ETH+DVOL certificati: 104 ipotesi distinte (11 famiglie), un agente-finder per ipotesi, verifica avversariale a 3 scettici sui promettenti, sintesi (153 agenti totali). Esito: NIENTE di nuovo regge -> conferma del soffitto strutturale ~1.3 BTC/ETH-direzionale; lo stack TP01+XS01+VRP01 resta imbattuto. - altlib.py: harness condiviso vettoriale leak-free (eval_weights/study_weights, fee-sweep, both-asset + hold-out 2025+). Riproduce i numeri canonici di TP01. - MARGINAL SCORER (study_marginal/marginal_vs_tp01): Sharpe INCREMENTALE vs baseline TP01 (corr, blend uplift OOS, alpha residua) + jackknife OOS (clean-year + drop-best-month). earns_slot = abs!=FAIL & ADDS & robust_oos. Smaschera gli overlay su TSMOM con PASS assoluti fasulli (CMB04, VOL11, ...) e il falso positivo KAMA (ADDS ma muore al jackknife). - runs/*.py (104) script riproducibili per ipotesi; wf_altstrat.js workflow. - Verdetto: 0 candidati deployabili; 2 LEAD fragili (VOL08, STA05_LS) da forward-monitor. - test_marginal_scorer.py blocca baseline + invarianti. Suite: 32 verde. Diario: docs/diary/2026-06-20-alt-strategies-100agent-sweep.md Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Adriano Dal Pastro
@@ -0,0 +1,187 @@
|
||||
"""SEA08 — US-session momentum on 1h bars.
|
||||
|
||||
HYPOTHESIS: On 1h: go long during 13-21 UTC when the prior (Asian+London) session
|
||||
was positive; otherwise flat. Idea: captures US risk-on drift when prior price
|
||||
action was constructive.
|
||||
|
||||
CAUSALITY CHECK:
|
||||
- "Prior session" = we look at the cumulative return of bars from the prior day's
|
||||
Asian+London window (00-12 UTC) that CLOSED before bar[i].
|
||||
- We compute the prior-session return as the log return from close[previous_day_00:00 UTC]
|
||||
to close[current_day_12:00 UTC], decided at bar[i] open (i.e., at close[i-1]).
|
||||
- Actually, we'll compute it simpler: the bar that ENDS at 12:00 UTC (the last
|
||||
Asian/London bar), and compare vs the bar that started the day (00:00 UTC).
|
||||
- For each hourly bar[i], at close[i-1] (= open of bar[i]), we know:
|
||||
* current UTC hour of bar[i]
|
||||
* the close at 12:00 UTC of today (if past 12:00)
|
||||
* the open at 00:00 UTC of today
|
||||
- Implementation: for each bar ending at time t (with UTC hour h):
|
||||
* If h in [13,21]: session is active
|
||||
* prior_session_return = (close at 12:00 of the current day / close at 00:00 of current day) - 1
|
||||
* We read close[i-1] with hour h (0-indexed, bar closes at h:00 UTC = bar represents h-1:00 to h:00)
|
||||
* Position at bar i = long (1.0) if h in [14..22] (bars DURING 13-21 UTC) AND prior session positive
|
||||
|
||||
Wait - let me be precise about 1h bar labeling:
|
||||
- A bar timestamped at "13:00 UTC" represents the candle from 12:00 to 13:00 UTC.
|
||||
- "close[13:00]" = price at end of 13:00 bar = price at 13:00 UTC.
|
||||
|
||||
For US session: we want to be long FROM 13:00 UTC TO 21:00 UTC.
|
||||
- We want to hold during bars whose close times are 14:00, 15:00, ..., 21:00 UTC
|
||||
(i.e., the bar from 13:00-14:00, ..., 20:00-21:00).
|
||||
|
||||
CAUSAL DECISION AT close[i]:
|
||||
- For each bar[i], we compute target[i] (what position to hold during bar i+1).
|
||||
- Bar i+1 closes at hour h+1.
|
||||
- We want to be long during bar i+1 if h+1 in {14,15,...,21}.
|
||||
- So target[i] = 1 if h in {13,...,20} AND prior_session_ret > 0.
|
||||
- prior_session_ret: from close at midnight (00:00 UTC) to close at noon (12:00 UTC) of the same day.
|
||||
- At close[i] with h in [13..20], we already know close[12:00] of today (it's in the past).
|
||||
|
||||
GRID: 3 variants tested to find best config:
|
||||
1. Pure time filter (no prior session condition)
|
||||
2. Prior session > 0 (baseline hypothesis)
|
||||
3. Prior session + vol-target scaling
|
||||
|
||||
We keep TF = 1h only (the hypothesis is inherently intraday on 1h bars).
|
||||
Total backtests: 1 tf × 3 variants × 2 assets = 6. Within budget.
|
||||
"""
|
||||
|
||||
import sys
|
||||
sys.path.insert(0, "/opt/docker/PythagorasGoal/scripts/research/alt")
|
||||
import altlib as al
|
||||
import numpy as np
|
||||
import pandas as pd
|
||||
|
||||
|
||||
def _build_session_features(df: pd.DataFrame):
|
||||
"""
|
||||
For each 1h bar at index i:
|
||||
- dt[i] = the UTC datetime when this bar closes (label of bar)
|
||||
- hour[i] = UTC hour of bar close
|
||||
- prior_session_ret[i] = return from close at 00:00 UTC to close at 12:00 UTC
|
||||
of the same day as bar[i], computed CAUSALLY (only available if bar[i] closes after 12:00 UTC).
|
||||
Returns (hour_arr, prior_session_ret_arr).
|
||||
"""
|
||||
dt = pd.to_datetime(df["datetime"], utc=True)
|
||||
close = df["close"].values.astype(float)
|
||||
n = len(df)
|
||||
|
||||
hour_arr = dt.dt.hour.values # UTC hour of bar close
|
||||
|
||||
# Build a lookup: for each (date, hour_target) -> close price
|
||||
# We need close at 00:00 UTC and close at 12:00 UTC for each date.
|
||||
#
|
||||
# The bar timestamped/labeled at 00:00 UTC closes at midnight = end of prior day.
|
||||
# So "open of day" price = close of the 23:00 bar (previous day) or close of 00:00 bar.
|
||||
#
|
||||
# Let's use simpler: close at 12:00 UTC bar (hour==12) as end of prior session.
|
||||
# Anchor = close at 00:00 UTC bar (hour==0) as start of day.
|
||||
# prior_session_ret = close[12:00] / close[00:00] - 1, for the same calendar date.
|
||||
#
|
||||
# To be causal at bar[i] with hour[i] >= 13: we need close[12:00] of same day,
|
||||
# which was available since 12:00 UTC (in the past).
|
||||
|
||||
# Build date -> index of 00:00 and 12:00 bars
|
||||
dates = dt.dt.date.values
|
||||
|
||||
# For each bar, find the closest prior-session data
|
||||
prior_ret = np.full(n, np.nan)
|
||||
|
||||
# Create a series indexed by datetime for easy lookup
|
||||
close_series = pd.Series(close, index=dt)
|
||||
|
||||
# Group by date to find the 00:00 and 12:00 anchors per day
|
||||
date_anchors = {} # date -> (close_00, close_12)
|
||||
|
||||
for i in range(n):
|
||||
d = dates[i]
|
||||
h = hour_arr[i]
|
||||
if d not in date_anchors:
|
||||
date_anchors[d] = [np.nan, np.nan] # [close_00, close_12]
|
||||
if h == 0:
|
||||
date_anchors[d][0] = close[i]
|
||||
elif h == 12:
|
||||
date_anchors[d][1] = close[i]
|
||||
|
||||
# Now fill prior_ret for each bar
|
||||
for i in range(n):
|
||||
d = dates[i]
|
||||
h = hour_arr[i]
|
||||
# Only compute if bar is in US session window and after 12:00 UTC
|
||||
if h >= 13 and d in date_anchors:
|
||||
c00, c12 = date_anchors[d]
|
||||
if np.isfinite(c00) and np.isfinite(c12) and c00 > 0:
|
||||
prior_ret[i] = c12 / c00 - 1.0
|
||||
|
||||
return hour_arr, prior_ret
|
||||
|
||||
|
||||
def target_time_only(df: pd.DataFrame) -> np.ndarray:
|
||||
"""
|
||||
Variant 1: Pure US-session time filter (13-21 UTC), no prior-session condition.
|
||||
Long during US session hours, flat otherwise.
|
||||
target[i] = 1.0 if bar[i+1] is in US session, else 0.0
|
||||
= 1.0 if hour[i] in {13,...,20} (so bar i+1 closes at 14..21 UTC).
|
||||
"""
|
||||
hour_arr, _ = _build_session_features(df)
|
||||
# target[i] = position held during bar i+1
|
||||
# bar i+1 closes at hour (hour_arr[i] + 1) % 24 approximately,
|
||||
# but let's use: hold long if hour[i] in 13..20 so we're long during 13:00->21:00 window
|
||||
target = np.where((hour_arr >= 13) & (hour_arr <= 20), 1.0, 0.0)
|
||||
return target
|
||||
|
||||
|
||||
def target_prior_session_momentum(df: pd.DataFrame) -> np.ndarray:
|
||||
"""
|
||||
Variant 2: Long during US session (13-21 UTC) ONLY IF prior session (00-12 UTC) was positive.
|
||||
"""
|
||||
hour_arr, prior_ret = _build_session_features(df)
|
||||
|
||||
# Propagate prior_ret within the US session of the same day
|
||||
# For bars in 13-21 UTC, prior_ret should already be set.
|
||||
# For continuity: once we set prior_ret at h=13, keep it for h=14..20 of same day.
|
||||
# Actually our loop sets it for all h>=13 of each day already.
|
||||
|
||||
us_session = (hour_arr >= 13) & (hour_arr <= 20)
|
||||
prior_positive = np.isfinite(prior_ret) & (prior_ret > 0)
|
||||
|
||||
target = np.where(us_session & prior_positive, 1.0, 0.0)
|
||||
return target
|
||||
|
||||
|
||||
def target_prior_session_vol_targeted(df: pd.DataFrame) -> np.ndarray:
|
||||
"""
|
||||
Variant 3: Like Variant 2 but with vol-targeting (20% annualized vol, cap 2x).
|
||||
"""
|
||||
direction = target_prior_session_momentum(df)
|
||||
return al.vol_target(direction, df, target_vol=0.20, vol_win_days=30, leverage_cap=2.0)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
print("SEA08 — US-session momentum on 1h bars")
|
||||
print("Testing 3 variants on 1h TF...")
|
||||
print()
|
||||
|
||||
# Variant 1: pure time filter
|
||||
rep1 = al.study_weights("SEA08-v1-time-only", target_time_only, tfs=("1h",))
|
||||
print(al.fmt(rep1))
|
||||
print()
|
||||
|
||||
# Variant 2: prior session momentum condition
|
||||
rep2 = al.study_weights("SEA08-v2-prior-session", target_prior_session_momentum, tfs=("1h",))
|
||||
print(al.fmt(rep2))
|
||||
print()
|
||||
|
||||
# Variant 3: vol-targeted version
|
||||
rep3 = al.study_weights("SEA08-v3-vol-target", target_prior_session_vol_targeted, tfs=("1h",))
|
||||
print(al.fmt(rep3))
|
||||
print()
|
||||
|
||||
# Pick the best config by holdout Sharpe
|
||||
reps = [rep1, rep2, rep3]
|
||||
best = max(reps, key=lambda r: r["verdict"].get("best_holdout_sharpe", -9))
|
||||
|
||||
print("=== BEST CONFIG ===")
|
||||
print(al.fmt(best))
|
||||
print()
|
||||
print("JSON:", al.as_json(best))
|
||||
Reference in New Issue
Block a user