PythagorasGoal/scripts/research/alt/runs/CMB03.py

"""CMB03 — Multi-TF trend confirm (4h fast + 1d slow agreement).

HYPOTHESIS: On the 4h TF, go long only when the 1d trend (TSMOM or SMA50)
agrees (is bullish). The intuition is that a fast-TF TSMOM signal might have
more noise; filtering by the slow TF trend reduces false signals.

CAUSAL ALIGNMENT (critical - see obs 4866):
  - 1d bar at timestamp T closes at end of day T. The 4h bar that CLOSES at
    the same time or later (within day T+1 onwards) can use it causally.
  - We compute the 1d signal on the 1d dataframe, then merge_asof onto 4h
    using the 1d bar CLOSE timestamp -> the 4h bar is valid only AFTER the
    1d bar has fully closed (direction="forward" with offset to avoid using
    the still-open 1d bar).
  - Implementation: for each 1d bar at timestamp T_close, the signal becomes
    available at T_close (the bar just closed). We map it to 4h bars whose
    open timestamp >= T_close (i.e. the NEXT 4h bar after the 1d bar closed).
    This means we use pandas merge_asof with left=4h open timestamps and
    right=1d close timestamps, direction="backward" — the 4h bar at open T
    gets the most recent 1d signal where 1d_close <= 4h_open.

GRID (4 configs x 2 assets x 1 TF = 8 backtests):
  A: 4h fast TSMOM (1m,3m) + 1d confirm SMA50 (price>SMA50)
  B: 4h fast TSMOM (1m,3m) + 1d confirm TSMOM (1m,3m,6m)
  C: 4h SMA crossover (20>50) + 1d confirm SMA50
  D: 4h SMA crossover (20>50) + 1d confirm TSMOM (1m,3m,6m)

All configs: long-only (0 or +1 direction), vol-targeted (20%, cap 2x).
"""

import sys
sys.path.insert(0, "/opt/docker/PythagorasGoal/scripts/research/alt")
import altlib as al
import numpy as np
import pandas as pd


# ---------------------------------------------------------------------------
# Helper: compute 1d trend signal and align causally to 4h bars
# ---------------------------------------------------------------------------

def _1d_tsmom_signal(df_1d: pd.DataFrame) -> np.ndarray:
    """TSMOM on 1d bars: long if majority of 1m/3m/6m horizons are positive.
    Returns array in {0, +1} (long-flat, no short).
    Decision at bar i uses close[i] (causal). Array indexed by 1d bar."""
    c = df_1d["close"].values.astype(float)
    bpd = al.bars_per_day(df_1d)  # should be ~1 for 1d
    horizons = [30 * bpd, 90 * bpd, 180 * bpd]
    votes = np.zeros(len(c))
    for h in horizons:
        h = int(h)
        sig = np.full(len(c), np.nan)
        if h < len(c):
            sig[h:] = np.sign(c[h:] / c[:-h] - 1.0)
        votes += np.nan_to_num(sig, nan=0.0)
    # Long when majority (>=1 out of 3) positive
    return np.where(votes > 0, 1.0, 0.0)


def _1d_sma50_signal(df_1d: pd.DataFrame) -> np.ndarray:
    """SMA50 trend on 1d: long when close > SMA50. Returns {0, +1}."""
    c = df_1d["close"].values.astype(float)
    sma50 = al.sma(c, 50)
    return np.where(c > sma50, 1.0, 0.0)


def _align_1d_to_4h(df_1d: pd.DataFrame, signal_1d: np.ndarray,
                    df_4h: pd.DataFrame) -> np.ndarray:
    """Map 1d signal onto 4h bars CAUSALLY.

    A 1d bar at timestamp T (which is the bar's OPEN time in ms) closes at
    T + 86400000ms. We expose the signal AFTER the 1d bar has fully closed,
    i.e. it's available to 4h bars whose open time >= T + 86400000ms (the
    start of the next day).

    Procedure:
    1. Build a series: (1d_close_timestamp, signal_1d)
       1d_close_ts = df_1d["timestamp"] + 86400000 (next bar open = this bar closed)
    2. For each 4h bar (open timestamp), take the most recent 1d signal
       where 1d_close_ts <= 4h_open_ts  (merge_asof backward).
    3. Forward-fill NaN (no signal yet = 0).
    """
    # 1d bar open timestamps + period offset = close timestamp = next 4h eligible
    # Compute 1d bar period in ms: use median diff of timestamps
    ts_1d = df_1d["timestamp"].values.astype(np.int64)
    diffs_1d = np.diff(ts_1d)
    period_ms = int(np.median(diffs_1d)) if len(diffs_1d) > 0 else 86_400_000

    # 1d_close_ts: the moment this 1d bar closed (= open of the NEXT bar)
    close_ts_1d = ts_1d + period_ms  # available after this timestamp

    right = pd.DataFrame({
        "close_ts": close_ts_1d,
        "sig": signal_1d.astype(float),
    }).sort_values("close_ts")

    ts_4h = df_4h["timestamp"].values.astype(np.int64)
    left = pd.DataFrame({"open_ts": ts_4h})

    merged = pd.merge_asof(
        left,
        right.rename(columns={"close_ts": "open_ts"}),
        on="open_ts",
        direction="backward",
    )
    out = merged["sig"].values.astype(float)
    # NaN = no 1d bar has closed yet -> be conservative, no position
    out = np.nan_to_num(out, nan=0.0)
    return out


# ---------------------------------------------------------------------------
# Fast-TF (4h) signals
# ---------------------------------------------------------------------------

def _4h_tsmom(df_4h: pd.DataFrame) -> np.ndarray:
    """TSMOM on 4h: long if 1m and 3m horizons agree (majority of 2)."""
    c = df_4h["close"].values.astype(float)
    bpd = al.bars_per_day(df_4h)  # ~6 for 4h
    h1m = int(30 * bpd)
    h3m = int(90 * bpd)
    votes = np.zeros(len(c))
    for h in [h1m, h3m]:
        sig = np.full(len(c), np.nan)
        if h < len(c):
            sig[h:] = np.sign(c[h:] / c[:-h] - 1.0)
        votes += np.nan_to_num(sig, nan=0.0)
    # Long when net positive (at least 1 of 2)
    return np.where(votes > 0, 1.0, 0.0)


def _4h_sma_cross(df_4h: pd.DataFrame, fast=20, slow=50) -> np.ndarray:
    """SMA crossover on 4h: long when SMA(fast) > SMA(slow)."""
    c = df_4h["close"].values.astype(float)
    sma_f = al.sma(c, fast)
    sma_s = al.sma(c, slow)
    return np.where(sma_f > sma_s, 1.0, 0.0)


# ---------------------------------------------------------------------------
# Combined target functions (4h TF, 1d confirm)
# ---------------------------------------------------------------------------

def make_target(asset: str, fast_type: str, slow_type: str):
    """Return a target_fn(df_4h) -> position array.

    Because altlib calls target_fn(df) with the chosen TF df, we fetch the
    1d df inside the closure (cached by altlib.get).
    """
    def target_fn(df_4h: pd.DataFrame) -> np.ndarray:
        # 1d dataframe for same asset (cached)
        df_1d = al.get(asset, "1d")

        # Compute 1d confirmation signal
        if slow_type == "sma50":
            sig_1d = _1d_sma50_signal(df_1d)
        elif slow_type == "tsmom":
            sig_1d = _1d_tsmom_signal(df_1d)
        else:
            raise ValueError(f"Unknown slow_type: {slow_type}")

        # Align 1d signal onto 4h bars (causal)
        confirm_4h = _align_1d_to_4h(df_1d, sig_1d, df_4h)

        # Compute 4h fast signal
        if fast_type == "tsmom":
            fast_4h = _4h_tsmom(df_4h)
        elif fast_type == "sma_cross":
            fast_4h = _4h_sma_cross(df_4h)
        else:
            raise ValueError(f"Unknown fast_type: {fast_type}")

        # Combined: long only when BOTH signals agree
        direction = np.where((fast_4h > 0) & (confirm_4h > 0), 1.0, 0.0)

        # Vol-target (20%, cap 2x)
        return al.vol_target(direction, df_4h, target_vol=0.20,
                             vol_win_days=30, leverage_cap=2.0)

    return target_fn


# ---------------------------------------------------------------------------
# Grid: 4 configs
# ---------------------------------------------------------------------------
CONFIGS = [
    dict(fast="tsmom",    slow="sma50",  label="tsmom4h_sma50_1d"),
    dict(fast="tsmom",    slow="tsmom",  label="tsmom4h_tsmom_1d"),
    dict(fast="sma_cross", slow="sma50", label="smacross4h_sma50_1d"),
    dict(fast="sma_cross", slow="tsmom", label="smacross4h_tsmom_1d"),
]

print("=== CMB03: Multi-TF trend confirm (4h fast + 1d slow) ===")
print(f"Grid: {len(CONFIGS)} configs x 2 assets x 1 TF = {len(CONFIGS)*2} backtests\n")

results = []
for cfg in CONFIGS:
    label = cfg["label"]
    fast = cfg["fast"]
    slow = cfg["slow"]

    # Build per-asset target functions
    # study_weights calls target_fn(df) for each asset, but we need to know
    # WHICH asset to fetch the 1d df for. We use a workaround: wrap in a
    # function that identifies the asset by calling al.get for BTC then ETH
    # and matching timestamps.
    #
    # Cleaner approach: run each asset separately and combine.
    # altlib.study_weights iterates assets internally, so we need target_fn(df)
    # to know the asset. We do this by checking df timestamps against cached dfs.

    def _target_fn(df_4h, _fast=fast, _slow=slow):
        # Identify asset by matching df timestamps to known cached dfs
        ts = df_4h["timestamp"].values[0]
        # Try BTC first, then ETH
        for _asset in ("BTC", "ETH"):
            try:
                _df_check = al.get(_asset, "4h")
                if _df_check["timestamp"].values[0] == ts:
                    return make_target(_asset, _fast, _slow)(df_4h)
            except Exception:
                pass
        # Fallback: try matching by length or first close
        c0 = df_4h["close"].values[0]
        for _asset in ("BTC", "ETH"):
            _df_check = al.get(_asset, "4h")
            if abs(_df_check["close"].values[0] - c0) / c0 < 0.01:
                return make_target(_asset, _fast, _slow)(df_4h)
        # Last resort
        return make_target("BTC", _fast, _slow)(df_4h)

    rep = al.study_weights(
        f"CMB03-{label}",
        _target_fn,
        tfs=("4h",),
    )
    print(al.fmt(rep))
    print(f"  JSON: {al.as_json(rep)}\n")
    results.append((rep, cfg))


# ---------------------------------------------------------------------------
# Pick best config by min_asset_holdout_sharpe
# ---------------------------------------------------------------------------
def best_holdout(item):
    rep = item[0]
    cells = rep.get("cells", [])
    if not cells:
        return -99.0
    return max(c.get("min_asset_holdout_sharpe", -99.0) for c in cells)

results.sort(key=best_holdout, reverse=True)
best_rep, best_cfg = results[0]

print("\n" + "=" * 60)
print(f"BEST CONFIG: {best_cfg['label']}")
print(al.fmt(best_rep))
print("JSON:", al.as_json(best_rep))