Overall top-100 hits
26.18
average across horizons
Predicted top-100 avg actual rank
225.43
lower is better
Rank IC
0.0823
picks the better of two stocks
52.6% of the time
IC reliability
13.70
t-stat;
IC positive 74.9% of months
not adjusted for overlapping windows
not adjusted for overlapping windows
Horizon Breakdown
top 100 only| Horizon | Average Top-100 Hits | Predicted Top-100 Avg Actual Rank | Rank IC |
|---|---|---|---|
| 3m | 25.58 | 230.79 | 0.0635 |
| 6m | 26.44 | 224.37 | 0.0901 |
| 12m | 26.50 | 221.13 | 0.0934 |
Run Note
test summaryHypothesis: eligibility is a structural axis no weight can reach. Every continuous variation on strategy_000109 failed -- ~9,000 rank combinations, every feature added at three weights, monotone and piecewise transforms, product and interaction terms, robustness weighting, row-wise fundamental aggregation, and a rank-space blend with strategy_000079 -- and the ONLY additions improving it in all 3 blocks and both halves were the raw data-coverage indicators, which are an artifact. A gate is different: the evaluator drops gated rows to the bottom outright, and because rank IC scores the whole eligible universe rather than only the top 100, banishing names that finish near the bottom raises IC in a way no continuous weight can. Change from strategy_000109: add eligibility_filter dropping the lowest 5% of from_low_252d each date (~25 of ~490 names); scoring untouched. Result IC 0.08232 (t 13.70, positive 74.9%) vs 109's 0.08077 -- NEW BEST IC ON THE BOARD, and better on every reported metric than 109: hits 26.176 vs 26.144, avg actual rank 225.43 vs 225.71. Paired t vs 109 is 2.10, winning 53.9% of 375 observations. IMPROVED but modestly. Honest caveat: gate size is NOT stable across time. Blocks 1-2 and both halves prefer a larger gate up to 7%, block 3 prefers smaller; at 5% this gains in b1/b2/h1/h2 but gives up 0.0026 in b3 (0.0881 vs 0.0907). A 1% gate improves all five sub-periods but wins only 44% of individual observations, so it is a few large wins rather than a consistent effect -- 5% is the more defensible choice. Also settled this iteration: the strategy_000079 rank blend reaching IC 0.0832 offline is entirely its coverage term reappearing; with fam_size removed, blending 079 in is strictly harmful at every weight (0.0779 vs 0.0808). The legacy eleven-factor family contributes nothing orthogonal. Next: fold the gate into the volatility variant to strengthen the all-round model.
Run Details
strategy_test_results #231| id | 231 |
|---|---|
| strategy_id | strategy_000112 |
| strategy_title | Robust Core With A Deep-Drawdown Gate |
| strategy_hash | a44d2191771664c006fb134d2eed81c12016eddb6b3d697146925a2c3a84c632 |
| evaluator_version | v3_top100_monthly |
| start_date | 2015-01-01 |
| end_date | 2025-06-01 |
| created_at | 2026-07-25 14:04:32 |
| updated_at | 2026-07-25 14:04:32 |
Sample Top-100 Lists
2020-03-31 / 6m / 27 hitsPredicted top 100
AAPL, ABBV, ABT, ADBE, AES, AFL, AIG, AMD, AMT, AMZN, ANET, AVGO, BKNG, BMY, BSX, CAG, CCL, CDNS, CF, CI, CMG, CNC, CNP, CRM, CSCO, DAL, DRI, DVN, DXC, DXCM, ETSY, EXC, EXPE, EXR, FANG, FITB, FTNT, FTV, GILD, GOOGL, HPE, HWM, INTC, ISRG, IVZ, KEY, KLAC, KMI, KO, KR, LLY, LNC, LRCX, LYV, MA, MCHP, MCO, MDT, MET, MGM, MO, MPC, MS, MSFT, NCLH, NEE, NEM, NFLX, NKE, NOW, NVDA, NWL, ODFL, OKE, ORCL, OXY, PLD, PPL, PRU, PYPL, QCOM, RCL, ROST, RTX, SPG, SPGI, SRE, SYF, SYK, SYY, TER, TJX, TSLA, UAL, URI, V, VRTX, VTR, WELL, WMB
Actual top 100
AAP, AAPL, ABMD, ADBE, ADSK, ALB, ALGN, AMD, AMP, AMZN, APD, APH, APTV, ATVI, AVGO, BBY, BWA, CARR, CDNS, CE, CHRW, CMG, CMI, CPRT, CRM, CTAS, CTLT, CTSH, DD, DE, DHI, DHR, DRI, DXCM, EBAY, EMN, ETSY, EXPE, FCX, FDX, GLW, GPS, HAL, HD, IDXX, IPGP, JCI, KMX, LEG, LEN, LH, LOW, MAS, MCHP, MGM, MOS, NCLH, NKE, NOW, NSC, NVDA, NWS, NWSA, ORLY, PAYC, PH, PHM, PNR, POOL, PVH, PWR, PYPL, QCOM, QRVO, RCL, ROK, ROL, SHW, SIVB, SNPS, SWK, SWKS, SYF, TDG, TEL, TER, TGT, TMO, TSCO, TSLA, TT, UPS, URI, VAR, VIAC, VTR, WHR, WMB, WST, WY
Matched tickers
AAPL, ADBE, AMD, AMZN, AVGO, CDNS, CMG, CRM, DRI, DXCM, ETSY, EXPE, MCHP, MGM, NCLH, NKE, NOW, NVDA, PYPL, QCOM, RCL, SYF, TER, TSLA, URI, VTR, WMB
Strategy Script
a44d219177"""Generated V3 stock-ranking strategy."""
from __future__ import annotations
import numpy as np
import pandas as pd
STRATEGY_ID = "strategy_000112"
STRATEGY_TITLE = "Robust Core With A Deep-Drawdown Gate"
DESCRIPTION = """
Every continuous variation on strategy_000109 has failed. Roughly nine thousand
rank combinations, every single feature added at three weights, monotone and
piecewise transforms, interaction and product terms, robustness weighting,
row-wise fundamental aggregation and a rank-space blend with strategy_000079
were all tested, and the only additions that improved it in all three blocks and
both halves were the raw data-coverage indicators, which are an artifact.
One structural axis was untested: eligibility. A gate is not a tilt. The
evaluator drops gated rows to the bottom of the ordering outright, and because
rank IC scores the entire eligible universe rather than only the top 100,
correctly banishing names that will finish near the bottom raises IC in a way no
continuous weight can.
The candidate is the deepest drawdown names. Being far above the 252-day low is
the strongest single signal on this objective at IC 0.0590, so the bottom of
that distribution is the part of the universe the model is most confident about.
The gate removes the lowest 5 percent of from_low_252d each date, which is about
25 names of roughly 490, and leaves the scoring untouched.
Gate size is not stable across time, and that is worth stating plainly: the
first two blocks and the first half prefer a larger gate, up to 7 percent, while
the third block prefers a smaller one. At 5 percent the model improves the first
block, the middle block, the first half and the second half but gives up 0.0026
in the third block. A 1 percent gate improves all five, but wins only 44 percent
of individual observations against 54 percent for this one, so it looks like a
handful of large wins rather than a consistent effect. The 5 percent version is
the honest choice.
"""
FACTORS_USED = [
"from_low_252d",
"volume",
"free_cash_flow_margin_pct",
"revenue_growth_pct"
]
PARAMETERS = {
"from_low_252d": 1.0,
"volume": 1.0,
"free_cash_flow_margin_pct": 1.0,
"revenue_growth_pct": 1.0
}
GATE_PERCENTILE = 0.05
COMPLEXITY = 5
PARENT_STRATEGY_ID = "strategy_000109"
GENERATION = 112
RANDOM_SEED = 20261120
CREATED_AT = "2026-07-25T00:00:00+00:00"
def _date_group(df):
if "trade_date" in df.columns:
return df["trade_date"]
return pd.Series(0, index=df.index)
def _rank_pct(df, column):
"""Date-local rank percentile centred on zero; missing values stay neutral."""
if column not in df.columns:
return pd.Series(0.0, index=df.index)
clean = df[column].replace([np.inf, -np.inf], np.nan)
ranked = clean.groupby(_date_group(df)).rank(pct=True)
return ranked.fillna(0.5) - 0.5
def eligibility_filter(df):
"""Drop the deepest 5 percent of drawdowns each date; the rest stay eligible."""
return _rank_pct(df, "from_low_252d") > (GATE_PERCENTILE - 0.5)
def calculate_score(df):
score = pd.Series(0.0, index=df.index)
for column, weight in PARAMETERS.items():
score = score + weight * _rank_pct(df, column)
total = sum(abs(w) for w in PARAMETERS.values()) or 1.0
score = score / total
return score.replace([np.inf, -np.inf], np.nan).fillna(0.0)