FACTOR / MOMENTUM-20D
First formal case · Dynamic study verifiedThe first complete study spanning point-in-time data boundaries, multi-horizon evidence, portfolio meaning, and reusable engineering components.
HYPOTHESIS
Test whether stocks with stronger trailing 20-day performance retain higher future-return ranks, and whether that relationship persists across 1, 5, 10, 20, and 60 days and through time. Twenty days is the lookback window, not a fixed holding period.
Twenty days is the lookback window; validation spans forward 1 / 5 / 10 / 20 / 60 days.
METHOD
IMPLEMENTATION
The critical order is factor first, point-in-time eligibility second; reversing it would truncate lookback history around index entry.
factor_raw = calculate_momentum_factor(
price_df, window=20
)
factor_eligible = factor_raw.where(eligible_mask)The eligibility mask controls daily ranking access; it does not define the factor.
01 / INFORMATION COEFFICIENT
One cross-sectional test per day: do factor ranks and future-return ranks move in the same direction?
02 / STABILITY
A 252-observation window approximates one trading year. Cross-horizon values are aligned to the common end date 2026-05-08.
03 / MONOTONICITY
Q1 is lowest momentum and Q5 highest. One day is mildly momentum-like; 5–60 days generally show an in-sample Q1-over-Q5 reversal.
The direction can change with the prediction horizon. Twenty days is only the historical lookback; forward one day and forward twenty days test different questions.
04 / TURNOVER
Membership turnover = 1 − overlap with the prior group / current group size. It is not traded-notional turnover, but it exposes cost sensitivity early.
05 / GROSS PORTFOLIOS
Rₚ,ₜ = Σwᵢ,ₜ₋₁Rᵢ,ₜ. Weights lag by one trading day; these are gross research portfolios without costs or fill constraints.
06 / ENGINEERING AFTER RESEARCH
The new components do not retrofit the first formal factor. They are independently verified and will be adopted from the next factor onward.
Top/Bottom percentile selection
Equal weight plus an extensible registry
Cadence, carry-forward, and execution lag
Strict weight/return alignment
Preserve factor one; adopt from factor two
Across the 20-year sample, old and new Q1/Q5 return series each contain 4,838 dates with zero maximum absolute difference. Top-20% ceil selection and quintile Q5 choose 54 versus 53 names when 266 are eligible; the 0.0185185 gap is definitional, not a bug.
EVIDENCE & LIMITS
The dynamic eligibility mask is formally integrated across 4,839 valid factor dates and 904 historical securities.
Rank IC means are negative at 1–60 days; 5–60-day quantiles generally show an in-sample Q1-over-Q5 reversal pattern.
The realized next-day Q5 gross portfolio outperforms Q1, consistent with the mildly positive one-day quantile spread; horizons answer different questions.
Reusable selection, weighting, and portfolio-return components were extracted and verified after the study; the next factor will adopt them.
All results are in-sample research evidence, not promised returns or a substitute for a cost-aware executable backtest.