You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
fithia2/tests/unit/backtest/test_cross_sectional_moment...

392 lines
15 KiB
Python

"""Unit tests for the CrossSectionalMomentum (12-1) engine.
Like every continuous-rotation engine on this codebase, the non-negotiable bar
is the look-ahead defense: every used bar date MUST be strictly before
decision_date. The momentum window AND the rebalance gate AND the universe
filters all run through `bar_provider.get_bars_before` which enforces this.
The Phase 19 pre-commit abort criteria are documented in the engine module
and the POC config (configs/experiments/xsmom_poc_v1.json). Tests here cover
pure logic, lookahead defense, and the rebalance-day semantics.
"""
from __future__ import annotations
import datetime as dt
from typing import Any
import pytest
from libs.backtest.domain import LookaheadViolationError, StrategyEngineConfig
from libs.backtest.cross_sectional_momentum import (
CROSS_SECTIONAL_MOMENTUM_EVENT_TYPE,
_SnapshotStoreBarAdapter,
build_candidates,
compute_12_1_momentum,
compute_20d_volatility,
compute_avg_dollar_volume_20d,
is_rebalance_day,
)
# ---------------------------------------------------------------------------
# Helpers
# ---------------------------------------------------------------------------
def _make_engine(**overrides: Any) -> StrategyEngineConfig:
base: dict[str, Any] = dict(
engine_id="xsmom_12_1_long",
event_types=[CROSS_SECTIONAL_MOMENTUM_EVENT_TYPE],
direction="long_only",
timing_class="after_close",
entry_timing_policy="next_open",
max_holding_days=21,
xsmom_enabled=True,
xsmom_lookback_days=60, # smaller for tests
xsmom_skip_days=5,
xsmom_top_n=3,
xsmom_holding_days=21,
xsmom_momentum_min=0.0,
xsmom_min_avg_dollar_volume=1_000_000.0,
xsmom_min_price=5.0,
xsmom_volatility_20d_max=0.10,
xsmom_stop_pct=0.10,
xsmom_target_pct=0.30,
)
base.update(overrides)
return StrategyEngineConfig(**base)
def _business_days(start: dt.date, count: int) -> list[dt.date]:
out: list[dt.date] = []
cursor = start
while len(out) < count:
if cursor.weekday() < 5:
out.append(cursor)
cursor = cursor + dt.timedelta(days=1)
return out
def _linear_bars(
days: list[dt.date],
*,
start_close: float = 100.0,
end_close: float = 130.0,
volume: float = 5_000_000.0,
) -> dict[dt.date, dict[str, Any]]:
"""Linear close path from start_close to end_close across all bars."""
n = len(days)
out: dict[dt.date, dict[str, Any]] = {}
for i, d in enumerate(days):
c = start_close + (end_close - start_close) * (i / max(1, n - 1))
out[d] = {
"open": c,
"high": c + 0.5,
"low": c - 0.5,
"close": c,
"volume": volume,
}
return out
# ---------------------------------------------------------------------------
# Pure trigger computations
# ---------------------------------------------------------------------------
def test_12_1_momentum_basic() -> None:
days = _business_days(dt.date(2024, 1, 2), 80) # need lookback+skip=65, plenty
bars_dict = _linear_bars(days, start_close=100.0, end_close=200.0)
bars = sorted(bars_dict.items(), key=lambda kv: kv[0])
# Use full window: lookback=60, skip=5 → window is bars[-65:-5]
momentum, used_dates = compute_12_1_momentum(bars, lookback_days=60, skip_days=5)
assert momentum is not None
# The window excludes the most recent 5 bars. With linear 100→200 across 80 bars,
# start of window is bars[-65] and end is bars[-6]. Both >100, <200.
start_close = bars[-65][1]["close"]
end_close = bars[-6][1]["close"]
expected = (end_close / start_close) - 1.0
assert abs(momentum - expected) < 1e-9
assert len(used_dates) == 60
def test_12_1_momentum_insufficient_history() -> None:
days = _business_days(dt.date(2024, 1, 2), 30)
bars_dict = _linear_bars(days)
bars = sorted(bars_dict.items(), key=lambda kv: kv[0])
momentum, used_dates = compute_12_1_momentum(bars, lookback_days=60, skip_days=5)
assert momentum is None
assert used_dates == []
def test_20d_volatility_basic() -> None:
days = _business_days(dt.date(2024, 1, 2), 30)
bars_dict = _linear_bars(days)
bars = sorted(bars_dict.items(), key=lambda kv: kv[0])
vol = compute_20d_volatility(bars)
assert vol is not None
assert vol >= 0
def test_avg_dollar_volume_20d_basic() -> None:
days = _business_days(dt.date(2024, 1, 2), 30)
bars_dict = _linear_bars(days, start_close=100, end_close=100, volume=1_000_000.0)
bars = sorted(bars_dict.items(), key=lambda kv: kv[0])
adv = compute_avg_dollar_volume_20d(bars)
# close ≈ 100, volume = 1M, so dollar volume ≈ 100M
assert abs(adv - 100_000_000.0) < 1_000_000.0
# ---------------------------------------------------------------------------
# Rebalance-day semantics
# ---------------------------------------------------------------------------
def test_is_rebalance_day_first_trading_day_of_new_month() -> None:
# Last bar is 2024-01-31 (Wed); decision_date is 2024-02-01 (Thu)
bars = [(dt.date(2024, 1, 31), {"close": 100.0})]
assert is_rebalance_day(bars, dt.date(2024, 2, 1)) is True
def test_is_rebalance_day_mid_month() -> None:
# Last bar is 2024-01-15 (Mon); decision_date 2024-01-16 (Tue) — same month
bars = [(dt.date(2024, 1, 15), {"close": 100.0})]
assert is_rebalance_day(bars, dt.date(2024, 1, 16)) is False
def test_is_rebalance_day_empty_bars_returns_false() -> None:
assert is_rebalance_day([], dt.date(2024, 1, 2)) is False
def test_is_rebalance_day_lookahead_raises() -> None:
bars = [(dt.date(2024, 2, 1), {"close": 100.0})]
with pytest.raises(LookaheadViolationError):
is_rebalance_day(bars, dt.date(2024, 2, 1))
# ---------------------------------------------------------------------------
# Lookahead defense: build_candidates rejects T+0 bars
# ---------------------------------------------------------------------------
class _LeakyProvider:
"""Returns bars INCLUDING the decision_date — must be caught by build_candidates."""
def __init__(self, bars_by_symbol: dict[str, dict[dt.date, dict[str, Any]]]) -> None:
self.bars_by_symbol = bars_by_symbol
def get_bars_before(
self, symbol: str, as_of_date: dt.date, lookback_days: int
) -> list[tuple[dt.date, dict[str, Any]]]:
# INTENTIONALLY LEAKY: returns bars where date == as_of_date too.
sym_bars = self.bars_by_symbol.get(symbol.upper(), {})
ordered = sorted(sym_bars.items(), key=lambda kv: kv[0])
return [(d, b) for d, b in ordered if d <= as_of_date][-lookback_days:]
def test_build_raises_lookahead_when_provider_returns_t0_bar() -> None:
"""Proof-by-contradiction: a leaky provider that returns T+0 must trip the assertion."""
days = _business_days(dt.date(2024, 1, 2), 80)
decision_date = days[-1] # final day is decision_date — leaky provider returns it
next_date = decision_date + dt.timedelta(days=1)
# Ensure decision_date is first trading day of new month — pick a setup that's a rebalance.
# Easier: just patch the rebalance check by using a non-rebalance date but with a leaky T+0;
# the leak should trip BEFORE rebalance gate via the explicit assertion.
bars_dict = _linear_bars(days)
leaky = _LeakyProvider({"FOO": bars_dict})
engine = _make_engine()
with pytest.raises(LookaheadViolationError):
build_candidates(
decision_date=decision_date,
next_trading_date=next_date,
universe_symbols=["FOO"],
engine=engine,
bar_provider=leaky,
)
def test_build_returns_empty_on_non_rebalance_day() -> None:
"""Mid-month decision_date must yield 0 candidates (no rebalance)."""
days = _business_days(dt.date(2024, 1, 2), 80)
bars_dict = _linear_bars(days)
bars_by_symbol = {f"SYM{i}": bars_dict for i in range(5)}
adapter = _SnapshotStoreBarAdapter(bars_by_symbol=bars_by_symbol)
engine = _make_engine()
# Use a mid-month date (where last_bar.month == decision_date.month).
decision_date = days[40] # somewhere mid-period
next_date = decision_date + dt.timedelta(days=1)
# Adjust: ensure last_bar (strictly before decision_date) is in same month
# The adapter strictly-before guarantees last_bar = days[39]. days[39] and days[40] same month.
candidates = build_candidates(
decision_date=decision_date,
next_trading_date=next_date,
universe_symbols=list(bars_by_symbol.keys()),
engine=engine,
bar_provider=adapter,
)
assert candidates == []
# ---------------------------------------------------------------------------
# Build candidates — happy path with multiple symbols + top-N selection
# ---------------------------------------------------------------------------
def test_build_emits_top_n_on_rebalance_day() -> None:
"""Construct 5 symbols with distinct momentum and verify top-3 emission."""
days = _business_days(dt.date(2024, 1, 2), 200)
# decision_date should be first trading day of a new month AND have
# >= lookback+skip bars (65) of history strictly before it.
rebalance_idx = None
for i in range(70, len(days)):
if days[i].month != days[i - 1].month:
rebalance_idx = i
break
assert rebalance_idx is not None
decision_date = days[rebalance_idx]
next_date = decision_date + dt.timedelta(days=1)
used_days = days[: rebalance_idx + 1] # bar_provider returns strictly-before, so up to days[rebalance_idx-1] are eligible
bars_by_symbol: dict[str, dict[dt.date, dict[str, Any]]] = {}
# 5 symbols with end_close in increasing order → momentum ranks ascending
end_closes = [110.0, 120.0, 130.0, 140.0, 150.0]
for i, end_c in enumerate(end_closes):
sym = f"SYM{i}"
bars_by_symbol[sym] = _linear_bars(used_days, start_close=100.0, end_close=end_c)
adapter = _SnapshotStoreBarAdapter(bars_by_symbol=bars_by_symbol)
engine = _make_engine(xsmom_top_n=3)
candidates = build_candidates(
decision_date=decision_date,
next_trading_date=next_date,
universe_symbols=list(bars_by_symbol.keys()),
engine=engine,
bar_provider=adapter,
)
assert len(candidates) == 3
# Highest momentum (SYM4) should be top-ranked
symbols_emitted = [c.symbol for c in candidates]
assert "SYM4" in symbols_emitted
assert "SYM3" in symbols_emitted
assert "SYM2" in symbols_emitted
assert "SYM0" not in symbols_emitted
assert "SYM1" not in symbols_emitted
def test_build_filters_negative_momentum() -> None:
"""Symbols with momentum < momentum_min must be dropped."""
days = _business_days(dt.date(2024, 1, 2), 200)
rebalance_idx = None
for i in range(70, len(days)):
if days[i].month != days[i - 1].month:
rebalance_idx = i
break
assert rebalance_idx is not None
decision_date = days[rebalance_idx]
next_date = decision_date + dt.timedelta(days=1)
used_days = days[: rebalance_idx + 1]
# 3 symbols with end_close BELOW start → negative momentum
bars_by_symbol = {
f"NEG{i}": _linear_bars(used_days, start_close=100.0, end_close=80.0 + i)
for i in range(3)
}
# 2 with positive
bars_by_symbol["POS1"] = _linear_bars(used_days, start_close=100.0, end_close=120.0)
bars_by_symbol["POS2"] = _linear_bars(used_days, start_close=100.0, end_close=140.0)
adapter = _SnapshotStoreBarAdapter(bars_by_symbol=bars_by_symbol)
engine = _make_engine(xsmom_top_n=10, xsmom_momentum_min=0.0)
candidates = build_candidates(
decision_date=decision_date,
next_trading_date=next_date,
universe_symbols=list(bars_by_symbol.keys()),
engine=engine,
bar_provider=adapter,
)
symbols_emitted = {c.symbol for c in candidates}
assert symbols_emitted == {"POS1", "POS2"}
def test_build_filters_volatility_max() -> None:
"""Symbols with 20d vol above the max must be dropped."""
days = _business_days(dt.date(2024, 1, 2), 200)
rebalance_idx = None
for i in range(70, len(days)):
if days[i].month != days[i - 1].month:
rebalance_idx = i
break
assert rebalance_idx is not None
decision_date = days[rebalance_idx]
next_date = decision_date + dt.timedelta(days=1)
used_days = days[: rebalance_idx + 1]
# SMOOTH: linear path → low vol
smooth_bars = _linear_bars(used_days, start_close=100.0, end_close=130.0)
# CHOPPY: alternating up/down by 20% from base, end at 130
choppy_bars: dict[dt.date, dict[str, Any]] = {}
for i, d in enumerate(used_days):
base = 100 + 30 * (i / max(1, len(used_days) - 1))
# Alternate ±20% on consecutive days
c = base * (1.2 if i % 2 == 0 else 0.8)
choppy_bars[d] = {"open": c, "high": c + 0.5, "low": c - 0.5, "close": c, "volume": 5_000_000.0}
bars_by_symbol = {"SMOOTH": smooth_bars, "CHOPPY": choppy_bars}
adapter = _SnapshotStoreBarAdapter(bars_by_symbol=bars_by_symbol)
engine = _make_engine(xsmom_top_n=10, xsmom_volatility_20d_max=0.05) # ~5% vol cap
candidates = build_candidates(
decision_date=decision_date,
next_trading_date=next_date,
universe_symbols=list(bars_by_symbol.keys()),
engine=engine,
bar_provider=adapter,
)
symbols_emitted = {c.symbol for c in candidates}
assert "CHOPPY" not in symbols_emitted
# SMOOTH may or may not survive depending on momentum_min — but it must NOT be the choppy one
# ---------------------------------------------------------------------------
# Engine disabled / no universe / etc — defensive
# ---------------------------------------------------------------------------
def test_build_returns_empty_when_engine_disabled() -> None:
engine = _make_engine(xsmom_enabled=False)
adapter = _SnapshotStoreBarAdapter(bars_by_symbol={})
candidates = build_candidates(
decision_date=dt.date(2024, 2, 1),
next_trading_date=dt.date(2024, 2, 2),
universe_symbols=["FOO"],
engine=engine,
bar_provider=adapter,
)
assert candidates == []
def test_build_raises_when_next_date_not_after_decision() -> None:
engine = _make_engine()
adapter = _SnapshotStoreBarAdapter(bars_by_symbol={})
with pytest.raises(LookaheadViolationError):
build_candidates(
decision_date=dt.date(2024, 2, 2),
next_trading_date=dt.date(2024, 2, 1), # before
universe_symbols=["FOO"],
engine=engine,
bar_provider=adapter,
)
def test_snapshot_store_adapter_returns_only_strictly_before() -> None:
"""The cached adapter must NEVER return bars where date >= as_of_date."""
days = [dt.date(2024, 1, d) for d in (2, 3, 4, 5, 8)] # mix of weekdays
bars = {d: {"close": 100.0 + i, "high": 101.0, "low": 99.0, "volume": 1e6} for i, d in enumerate(days)}
adapter = _SnapshotStoreBarAdapter(bars_by_symbol={"FOO": bars})
# as_of_date = days[3]; should return only days[:3]
out = adapter.get_bars_before("FOO", days[3], lookback_days=10)
out_dates = [d for d, _ in out]
assert all(d < days[3] for d in out_dates)
assert out_dates == days[:3]