@tangle-network/agent-bench 0.3.7 → 0.4.0

package/src/quant-arena/python/vbt-worker.py

@@ -0,0 +1,192 @@
+"""QUANT-ARENA vectorbt scoring worker.
+
+Persistent process speaking JSON-lines over stdin/stdout. The TypeScript side
+(vbt-client.ts) spawns ONE worker per campaign and streams score requests to
+it; the worker replies with per-window stats plus the full-range equity curve.
+
+Protocol (one JSON object per line):
+  request  {"id": n, "op": "ping"}
+  reply    {"id": n, "ok": true, "vbt": "<version>"}
+
+  request  {"id": n, "op": "score",
+            "open_prices":  [[T x N floats]],   # fill prices (next-open model)
+            "close_prices": [[T x N floats]],   # valuation prices (close-marked equity)
+            "weights":      [[N floats] | null, ... T rows],
+                                                # decision at day t; null = no rebalance
+                                                # (JSON has no NaN — null plays that role)
+            "cost_bps": 10, "slippage_bps": 5,
+            "windows": [[start, end], ...]}     # [start, end) day-index ranges
+  reply    {"id": n, "ok": true,
+            "windows": [{"totalReturn", "maxDD", "sharpe", "trades"}, ...],
+            "full":    {same shape, over [0, T)},
+            "equity":  [T floats]}              # group value, close-marked
+  errors   {"id": n, "ok": false, "error": "<message>"}
+
+Execution model (aligned with the TS reference engine, backtest.ts):
+  - a decision at the close of day t fills at day t+1's open: the worker
+    shifts the weights matrix down one row before handing it to vectorbt;
+  - Portfolio.from_orders(size_type='targetpercent', group_by=True,
+    cash_sharing=True, call_seq='auto'), fill price = open, valuation
+    (val_price for sizing) = open — matching the TS engine's "target dollars
+    are a fraction of equity marked at the fill's open";
+  - the TS engine charges slippage as a bps FEE on traded dollars, so the
+    worker folds slippage_bps into vectorbt's `fees` and passes slippage=0
+    (vectorbt's own slippage perturbs the fill price — a different model);
+  - stats replicate backtest.ts statsForRange exactly: sample-std (ddof=1)
+    annualized Sharpe at 252 days, positive-convention max drawdown with the
+    peak reset at each window start, trades counted on fill days inside
+    (start, end).
+
+Determinism: NUMBA_NUM_THREADS=1 is pinned before import so JIT-compiled
+reductions run single-threaded; a warmup call at startup absorbs the numba
+cold-JIT cost before the first real request.
+
+License note: vectorbt is Apache-2.0 + Commons Clause. This worker is for
+internal research use (scoring candidate strategies in the lab); it is not
+sold or offered as a commercial service.
+"""
+
+from __future__ import annotations
+
+import json
+import math
+import os
+import sys
+
+os.environ.setdefault("NUMBA_NUM_THREADS", "1")
+
+import numpy as np  # noqa: E402
+import pandas as pd  # noqa: E402
+import vectorbt as vbt  # noqa: E402
+
+TRADING_DAYS_PER_YEAR = 252
+
+
+def range_stats(equity: np.ndarray, order_days: np.ndarray, start: int, end: int) -> dict:
+    """Mirror of backtest.ts statsForRange over a value curve."""
+    if start < 0 or end > equity.shape[0] or end - start < 2:
+        raise ValueError(f"bad range [{start}, {end}) over {equity.shape[0]} days")
+    eq = equity[start:end]
+    total_return = float(eq[-1] / eq[0] - 1.0)
+    peak = -math.inf
+    max_dd = 0.0
+    for value in eq:
+        if value > peak:
+            peak = value
+        dd = (peak - value) / peak
+        if dd > max_dd:
+            max_dd = dd
+    returns = eq[1:] / eq[:-1] - 1.0
+    n = returns.shape[0]
+    mean = float(returns.mean())
+    if n > 1:
+        std = float(returns.std(ddof=1))
+    else:
+        std = 0.0
+    sharpe = (mean / std) * math.sqrt(TRADING_DAYS_PER_YEAR) if std > 0 else 0.0
+    trades = int(((order_days > start) & (order_days < end)).sum())
+    return {
+        "totalReturn": total_return,
+        "maxDD": float(max_dd),
+        "sharpe": float(sharpe),
+        "trades": trades,
+    }
+
+
+def run_portfolio(req: dict) -> tuple[np.ndarray, np.ndarray]:
+    """Returns (equity curve close-marked, order fill-day indices)."""
+    opens = np.asarray(req["open_prices"], dtype=np.float64)
+    closes = np.asarray(req["close_prices"], dtype=np.float64)
+    if opens.shape != closes.shape or opens.ndim != 2:
+        raise ValueError(f"price shape mismatch: open {opens.shape} vs close {closes.shape}")
+    T, N = opens.shape
+    weights_rows = req["weights"]
+    if len(weights_rows) != T:
+        raise ValueError(f"weights has {len(weights_rows)} rows for {T} days")
+
+    # Decision at close of t -> order row t+1, filled at t+1's open.
+    size = np.full((T, N), np.nan)
+    for t, row in enumerate(weights_rows):
+        if row is None:
+            continue
+        if t + 1 >= T:
+            continue  # a decision on the final bar can never fill
+        if len(row) != N:
+            raise ValueError(f"decision at t={t} has {len(row)} weights for {N} assets")
+        size[t + 1] = row
+
+    fee_rate = (float(req["cost_bps"]) + float(req["slippage_bps"])) / 10_000.0
+    columns = [f"A{k}" for k in range(N)]
+    close_df = pd.DataFrame(closes, columns=columns)
+    open_df = pd.DataFrame(opens, columns=columns)
+    size_df = pd.DataFrame(size, columns=columns)
+
+    pf = vbt.Portfolio.from_orders(
+        close=close_df,
+        size=size_df,
+        price=open_df,
+        val_price=open_df,
+        size_type="targetpercent",
+        group_by=True,
+        cash_sharing=True,
+        call_seq="auto",
+        fees=fee_rate,
+        slippage=0.0,
+        init_cash=1.0,
+        freq="1D",
+    )
+    equity = np.asarray(pf.value().values, dtype=np.float64)
+    order_days = np.asarray(pf.orders.records["idx"], dtype=np.int64)
+    return equity, order_days
+
+
+def score(req: dict) -> dict:
+    equity, order_days = run_portfolio(req)
+    T = equity.shape[0]
+    windows = [range_stats(equity, order_days, int(w[0]), int(w[1])) for w in req["windows"]]
+    return {
+        "windows": windows,
+        "full": range_stats(equity, order_days, 0, T),
+        "equity": [float(v) for v in equity],
+    }
+
+
+def warmup() -> None:
+    req = {
+        "open_prices": [[100.0, 50.0]] * 8,
+        "close_prices": [[101.0, 49.5]] * 8,
+        "weights": [None, [0.5, 0.4], None, None, [0.2, 0.1], None, None, None],
+        "cost_bps": 10,
+        "slippage_bps": 5,
+        "windows": [[0, 8]],
+    }
+    score(req)
+
+
+def main() -> None:
+    warmup()
+    sys.stdout.write(json.dumps({"ready": True, "vbt": vbt.__version__}) + "\n")
+    sys.stdout.flush()
+    for line in sys.stdin:
+        line = line.strip()
+        if not line:
+            continue
+        rid = None
+        try:
+            req = json.loads(line)
+            rid = req.get("id")
+            op = req.get("op")
+            if op == "ping":
+                reply = {"id": rid, "ok": True, "vbt": vbt.__version__}
+            elif op == "score":
+                reply = {"id": rid, "ok": True, **score(req)}
+            else:
+                reply = {"id": rid, "ok": False, "error": f"unknown op {op!r}"}
+        except Exception as exc:  # fail per-request, keep serving
+            reply = {"id": rid, "ok": False, "error": f"{type(exc).__name__}: {exc}"}
+        sys.stdout.write(json.dumps(reply) + "\n")
+        sys.stdout.flush()
+
+
+if __name__ == "__main__":
+    main()