price-model

A point-in-time (PIT) corrected cross-sectional equity return predictor on the S&P 500, evaluated under strict HP-free held-out protocols at the 21-day (approximately monthly) forward horizon.

Executive summary

TL;DR: A machine learning project proving that a simple 6-feature linear regression beats complex tree models at predicting 21-day stock returns in the current bull market.

A PIT-corrected, held-out 21-day cross-sectional return predictor on the S&P 500. The headline is reported regime-confined: trained only on the post-2022-10 bull regime (train 2022-10-10 → 2024-11-30, single refit; trees held-out-Optuna-tuned with the cutoff at 2024-12-31) and tested on 2025-01-02 → 2026-06-24 (348 dates). This is the honest "deploy when the regime resembles recent training" scenario; the expansion on longer time-scale across-regime is to be studied.

A 6-feature cross-sectional OLS regression on documented academic anomalies (Han-He-Rapach-Zhou 2024) is the strongest model on both gross signal and net-of-cost return — beating long-horizon momentum and held-out-tuned LightGBM / XGBoost / CatBoost:

Long-short Sharpe is for a top-20% / bottom-20% (quintile), equal-weighted portfolio, computed on the ~17 non-overlapping 21-day rebalances of the test window; net columns deduct round-trip turnover × cost (3 / 10 / 20 bp per side). Gross IC is the per-date rank correlation averaged over all OOS dates, but its t-stat is Newey–West HAC-corrected (Bartlett lag 21) to account for the 21-day return overlap — the naive daily t (≈ 10) is inflated ~3.7× by that overlap (realized deflation +10.1 → +2.72; √21 ≈ 4.6× is the perfect-overlap upper bound) and is not the honest figure. See Metric definitions.

Model (regime-confined)	Gross IC	t-stat (HAC)	Gross Sharpe	Net 3 bp	Net 10 bp	Net 20 bp	Annual turnover
OLS, 6-feature (Ridge ≡ Lasso)	+0.0871	+2.72	+2.09	+2.05	+1.97	+1.85	9×
36-month momentum factor	+0.0497	+2.31	+1.41	+1.39	+1.34	+1.28	4×
LightGBM, held-out Optuna (rank panel)	+0.0689	+1.64	+1.06	+1.02	+0.93	+0.79	11×
CatBoost, held-out Optuna (rank panel)	+0.0610	+1.30	+0.91	+0.88	+0.81	+0.71	10×
XGBoost, held-out Optuna (rank panel)	+0.0586	+1.29	+0.80	+0.76	+0.69	+0.58	10×

Only OLS (HAC t +2.72) and momentum (+2.31) are individually significant; none of the tuned trees clears |t| > 2 once the overlap is corrected (+1.3 to +1.6), so their apparent edge is not distinguishable from zero on this window.

Observations:

• OLS wins gross and net — the only configuration in this project where the model beats momentum net-of-cost, and by a wide margin: gross Sharpe +2.09 and, because the curated 6-feature book turns over only ~9×/yr, net +1.85 at 20 bp (vs momentum +1.28 and the best tree +0.79).
• Regularization is inert — the model is plain OLS. On this panel the IC-scored forward-chain CV drives both Lasso and Ridge to α→0; all three (OLS ≡ Ridge ≡ Lasso) produce identical predictions to every digit. With ~238k training rows × 6 features the Gram matrix is hugely over-determined, so shrinkage has nothing to do — the CV-IC curve decays monotonically with α. The story is the curated panel, not the regularizer. (Pruning the prior 9-feature panel to 6 — dropping idio_vol_20, max_return_21d, and the anti-generalizing beta_60 — is most of the lift.)
• Every tuned tree trails — and none is statistically significant. Even with held-out Optuna on each tree's best panel (all three prefer the 9-feature rank panel over the 14-feature engineered one), the trees net +0.6–0.8 at 20 bp, and once the 21-day overlap is HAC-corrected their IC t-stats fall to +1.3–1.6 — not distinguishable from zero. The signal is near-linear; trees over-parameterize it.
• The honest t-stat is +2.72, not ~+10. The naive daily IC t treats 348 overlapping-window ICs as independent; the 21-day overlap means there are really only ~17 independent monthly observations. The HAC correction (lag 21) deflates the OLS t from +10.1 to +2.72 — still significant at p < 0.01, but a far cry from the inflated figure.

For the technical setup (PIT correction, regime-confined walk-forward with embargo, the unified temporal IC-scored CV), see Methodology; for why OLS beats trees, why the edge is regime-bound, why regularization is inert, and the net-of-cost analysis, see Discussion.

The result is statistically rigorous, regime-conditional, and not deployable for retail investors after bid-ask spreads, commissions, and capital-gains taxes. See Data quality and methodological limitations and Scope and limitations for the bounds on what these numbers mean. For the precise definitions of each metric in the table above (IC, t-stat, gross / net long-short Sharpe, annual turnover), see Metric definitions.

Metric definitions

The headline 21-day result is reported in terms of standard cross-sectional asset-pricing metrics, defined below; the same formulas apply at any horizon.

• Information Coefficient (IC). Cross-sectional ranking quality of the model, averaged over time.

  • Per-date IC: The Spearman rank correlation between predictions and realized H-day forward excess returns for all N_t valid tickers on a given day t.
  • Time-averaged IC: The unweighted average of per-date ICs across all evaluation dates.
  • The Overlap Problem (Worked Example): While the 21-day target yields ~348 daily correlations, consecutive days share 20 of their 21 forward days. Thus, daily ICs are heavily autocorrelated. There are really only ~17 independent monthly observations, so a naive t-stat (+10.1) overstates significance ~3.7× (it deflates to +2.72; √21 ≈ 4.6× is the perfect-overlap ceiling).
  • Benchmarks: IC = +0.02 is a credible edge; +0.10 is exceptional; > +0.20 is generally unrealistic.

• t-stat of IC (HAC Corrected). Because the 21-day forward windows overlap, adjacent ICs are positively autocorrelated.

  • The reported t-stat applies a Newey–West HAC correction (Bartlett kernel, lag 21).
  • Formula: t_HAC = mean / √Var_HAC where Var_HAC = (1/n)[γ₀ + 2·Σ_{k=1}^{21}(1 − k/22)·γ_k].
  • For the 6-feature panel, this deflates the t-stat from +10.1 (naive) to +2.72 (HAC) — which is the honest, statistically significant figure (p < 0.01). (Equivalently, computing IC on the ~17 non-overlapping rebalance dates gives a comparable t ≈ 2.8.)

• Long-short Sharpe. Annualized Sharpe ratio of a portfolio that is long the top-quintile predicted tickers (top 20%) and short the bottom-quintile (bottom 20%), equal-weighted within each leg, rebalanced every 21 trading days and held to expiry — the deployable monthly schedule. The quintile cut is top_frac=0.2 (src/price_model/eval/metrics.py). Over the test window this is the ~17 non-overlapping 21-day long-short returns mean(top.realized) − mean(bottom.realized), with Sharpe = mean / stdev × √(252 / 21). Above +1.0 is the conventional bar for institutional tradeability; values are gross of cost unless labeled "net N bp."

  Rebalancing & assumptions. (i) dollar-neutral, equal-weight legs; (ii) the book is rebalanced once per 21-day horizon and the return series is non-overlapping, so — unlike a daily-formed overlapping estimator — there is no autocorrelation inflation of the Sharpe (the honest trade-off is a small ~17-observation sample); (iii) no financing, short-borrow, or capital-gains tax; (iv) annualization by √(252/21); (v) the net columns deduct round-trip turnover × cost on the same 21-day clock (see Annual turnover).

• Annual turnover. One-sided fraction of portfolio capital traded per year. Because the prediction targets a 21-day horizon, the book is held over the horizon and turnover is measured on the rebalance clock: mean(0.5 · Σ_i |w_t[i] − w_{t-21}[i]|) · (252/21), where w_t is the long-short weight vector and the t-21 lag matches the 21-day holding period (≈ 12 rebalances/year). An annual turnover of 1.0 means the portfolio replaces itself once per year; 9x means the equivalent of nine full-portfolio rotations per year. (A previous version measured a 1-day lag and annualized by 252 — charging daily churn against a 21-day return, which over-stated turnover ~21× and is now fixed.) The metric is implemented in src/price_model/eval/turnover.py and governs how much of a gross signal survives net of transaction costs: for cost level b bp per side, annual cost drag ≈ annual_turnover · b · 2 / 10000.

• After-cost Sharpe at N bp. Long-short Sharpe computed on returns net of transaction-cost drag. For each rebalance, gross return is mean(top.realized) − mean(bottom.realized); net return subtracts turnover_t · N / 10000 · 2 from gross (the · 2 charges the round trip — entering and later exiting each position). The resulting net-return series is annualized to a Sharpe using the same √(252/horizon) convention as gross Sharpe. Reported at 3, 10, and 20 bp per side to span institutional → retail cost regimes.

Reproduction

Step 0 — install and build the PIT universe (one-time, ~10-15 min cold)

# Core: required for everything below.
pip install -e ".[dev,classical]"

# Optional: notebook visualization stack (matplotlib + seaborn + jupyter).
# Install if you plan to run the notebooks/ directory, including the
# feature-exploration boilerplate at notebooks/05_feature_exploration.ipynb.
pip install -e ".[notebooks]"

# Scrape Wikipedia for historical S&P 500 membership and write the universe file
python -m price_model.cli build-universe --name sp500_pit --start 2017-01-01

# Fetch yfinance data for the ~700 resolving tickers (~10-15 min cold; subsequent
# runs are incremental, fetching only the new rows since the cache's max date)
python -m price_model.cli refresh-data --universe sp500_pit --start 2017-01-01

Step 1 — run the regime-confined headline

One script runs the whole comparison: held-out Optuna sweeps for the three trees on both panels (rank-9 and engineered-14), then trains the tuned trees + the 6-feature linear models (OLS, plus the equivalent Lasso/Ridge — identical on this panel) + the momentum factors. Everything is regime-confined — trained only on 2022-10-10 → 2024-11-30 (single refit) and tested on 2025+.

# The run writes to the single-writer DuckDB store; close any Jupyter kernel
# that opened it first. ~1.5-2h at 40 trials/model (CatBoost dominates).
bash scripts/run_headline_comparison.sh 40

Regime confinement is pure config — three walk_forward fields, no code edits:

• data.start: 2017-01-01 — full price history so the ~3-year momentum_756 warmup is satisfied.
• train_start: 2022-10-10 — training lower bound, applied identically across panels regardless of each panel's warmup (this is why the harness gained a train_start parameter; a single data.start warms the rank and engineered panels at different dates).
• first_refit: 2025-01-02 + refit_freq_days: 9999 — one train/test boundary; embargo_days: 33 (calendar) covers the 21-trading-day target.

The Optuna sweeps read the same warmed matrix, so they inherit train_start (CV folds are in-regime) and hold out 2025+ via --max-date 2024-12-31.

Step 2 — the headline net-of-cost table

PYTHONPATH=src python scripts/net_cost_regime_21d.py

Fits the linear family fresh (OLS / Ridge / Lasso on curated-6 with the temporal IC-scored CV — the store holds stale pre-fix predictions and no OLS) and reads the trees + momentum from the store, joins realized 21-day excess returns, and prints the IC point estimate (all dates) with its HAC-corrected t-stat (Newey–West, lag 21) plus gross / 3 / 10 / 20 bp net Sharpe on the non-overlapping 21-day rebalance schedule — the headline convention. (scripts/net_cost_regime.py reports the same on the daily-overlap estimator.) Use these, not scripts/compare_net_of_cost.py, for the regime headline: the shared compare scripts intersect the full accumulated store and drop most of the new models.

Step 3 (optional) — investigation diagnostics

The Discussion's claims reproduce via:

PYTHONPATH=src python scripts/inspect_alpha_selection.py    # prints CV folds, α grid, selected α: regularization → α≈0 (OLS)
PYTHONPATH=src python scripts/spectrum_comparison.py         # model × panel grid: OLS≡Ridge≡Lasso, linear>trees, panel dominates
PYTHONPATH=src python scripts/vol_ablation_lasso_ridge.py    # 9→6 prune; beta_60 is anti-generalizing
PYTHONPATH=src python scripts/inspect_headline_lasso_coefficients.py  # vol-cluster signs are signal, not L1 cancellation (Ridge agrees)
PYTHONPATH=src python scripts/survivorship_ablation_headline.py       # PIT-on vs survivor-snapshot, in-regime
# Deflated Sharpe (scripts/deflated_sharpe_audit.py) is parked for now — not used in the headline.

Methodology

The headline result is produced by a walk-forward, cross-validated, point-in-time-corrected backtest. This section documents each component.

Universe: S&P 500 with point-in-time membership filter

Point-in-time (PIT) correction restricts the cross-section evaluated on each date t to tickers that were actually members of the S&P 500 index on date t — not the tickers that are members today. The correction defends against survivorship bias: today's S&P 500 list disproportionately contains companies that survived, by selection. A backtest evaluated on today's list silently assumes the model would have "known" in 2018 to focus on the 2026 survivors.

The mechanics:

• src/price_model/data/sources/sp500_membership.py scrapes the Wikipedia "List of S&P 500 companies" page and its "Selected changes" table to reconstruct, for every ticker that has ever been in the index between 2014 and the present, an (added_date, removed_date) membership window.
• src/price_model/data/membership.py::filter_panel_to_pit(panel) drops every row whose (date, ticker) pair falls outside that ticker's membership window. A stock added on 2020-06-22 contributes only its post-2020-06-22 rows; a stock removed on 2019-04-30 contributes only its pre-2019-04-30 rows.

The PIT correction in this project is partial: yfinance does not return data for ~12% of historical S&P 500 tickers (SIVB, FRC, ATVI, AGN, etc. — see Data quality and methodological limitations). A bias-free PIT analysis requires a paid feed (CRSP or Norgate Premium Data).

Forecast target and walk-forward training

The forecast target is the cross-sectional excess log-return over a fixed horizon. For each date t and ticker i, the target y[t, i] is the log-return of i from t+1 to t+H minus the cross-sectional mean log-return over the same window. The headline uses H = 21 trading days (approximately one month); the same framework is implemented at H = 5 and is configurable per experiment via target_horizon in the YAML configs.

Regime-confined training (this branch). The headline trains on a single regime — the post-2022-10 bull regime — and tests on 2025+. This is one train/test block, controlled by three walk_forward config fields:

• data.start: 2017-01-01 — full price history, so the ~3-year momentum_756 warmup is satisfied.
• train_start: 2022-10-10 — the training lower bound. Applied identically across feature panels regardless of each panel's warmup length (the rank panel warms at 2022-10, the 14-feature engineered panel at 2021-10; without train_start a single data.start would train them on different windows). This is the regime boundary from the time-split in notebooks/03.
• first_refit: 2025-01-02 + refit_freq_days: 9999 — a single refit, so the training block is [2022-10-10, 2024-11-30] and the test slice is [2025-01-02, 2026-06-24] (348 dates).
• embargo_days: 33 (calendar) — covers the 21-trading-day (~31 calendar) target, so the last training label resolves before the test window. (This fixes the earlier embargo_days: 22 calendar-vs-trading unit bug; train_end = first_refit − 33 = 2024-11-30.)

train_start and first_refit are additive harness parameters (default None = the original expanding-window behavior, which trains from warmup-end across all regimes — still used on main). For that across-regime framing, omit them and use min_train_days/refit_freq_days as before.

Feature lookahead safety is verified by the truncation-invariance test in tests/test_no_leakage.py: each registered feature's value at date t must be unchanged when the panel is truncated to dates ≤ t. The test is parametrized over FEATURE_REGISTRY and runs in CI. Note it covers features only — target leakage across the train/test boundary is the embargo's job (now airtight at 33 days), not this test's.

Inner cross-validation for regularization-strength selection

Any regularization strength α (for Ridge / Lasso / ElasticNet) is selected by one unified grader, identical to the tree sweeps: a temporal, purged, expanding-window forward-chain CV scored on per-date IC (src/price_model/models/_cv.py). Within each refit's training block:

• The training dates are cut into cv+1 contiguous chunks (default cv=3 → 4 chunks); chunk i is a validation fold and all earlier dates form its training set, minus a cv_embargo-date purge (≥ the 21-day horizon) so the forward target cannot straddle the boundary. Expanding window, strictly causal — no future dates leak into α-selection.
• For each candidate α the model is fit on each fold's train rows and scored by mean per-date Spearman IC on its validation rows; the α maximizing the mean fold-IC is chosen, then the model is refit once on the full block.
• Grids are data-driven: Lasso/ElasticNet anchor to alpha_max = max_j|x_j·ȳ|/n (the α that zeros all coefficients); Ridge anchors to the Gram-matrix eigenvalue scale trace(XᵀX)/p (L2 has no alpha_max).

On the curated 6-feature panel this selection drives α → 0 — the model is effectively OLS. With ~238k training rows × 6 features the design is hugely over-determined, so shrinkage doesn't help: the CV-IC curve is flat at the low-α (OLS) end and decays monotonically as α bites. OLS, Lasso, and Ridge therefore produce identical predictions, and the headline is reported as OLS.

(Correction to an earlier version: α used to be selected by sklearn's LassoCV/RidgeCV, which on an integer cv build a non-temporal KFold(shuffle=False). Because the panel is sorted ["ticker","date"], those folds fell on ticker boundaries — a by-ticker, MSE-scored, fully-date-overlapping split that leaked future dates into HP selection. That was an accidental sklearn default, not a design choice; it is now replaced by the temporal IC grader above, so linear and tree hyperparameters are chosen the same temporally-honest way.)

The walk-forward outer loop independently guarantees each refit's α is selected only on data preceding the deployment slice.

Held-out Optuna protocol for LightGBM hyperparameter search

LightGBM has more hyperparameters than CV can search inside a refit. Instead, the project uses Optuna with purged walk-forward CV (de Prado, Advances in Financial Machine Learning, Ch. 7) for hyperparameter selection.

The HP-selection bias correction is provided by the --max-date flag in scripts/optuna_sweep.py: dates after the specified cutoff are excluded from the matrix before CV folds are constructed, so the chosen HPs are provably independent of any data in the post-cutoff evaluation slice. With train_start set, the sweep also confines its CV folds to the regime (lower bound), so HP selection is in-regime and held out from 2025+ via --max-date 2024-12-31.

Regime-confined sweeps. Each tree is swept on both panels (held-out, 2024-12-31 cutoff; CV = mean per-date IC over 3 purged forward-chain folds). The CV winner picks the deployed panel:

Model	Panel	CV mean IC	2025+ OOS IC
LightGBM	rank-9	+0.0506	+0.0685
LightGBM	eng-14	+0.0361	+0.0546
XGBoost	rank-9	+0.0502	+0.0583
XGBoost	eng-14	+0.0308	+0.0430
CatBoost	rank-9	+0.0513	+0.0606
CatBoost	eng-14	+0.0296	+0.0499

• CV predicts OOS. Every model's 2025+ OOS IC exceeds its CV IC (e.g. LightGBM 0.051 → 0.069) — no fragility penalty, because the CV folds and the test slice share the regime. The rank-9 panel wins CV for all three trees, so the panel choice is made honestly (on CV), not by peeking at OOS.

Headline feature panel: 6 curated anomalies (pruned from 9)

The headline uses 6 features — the 9-feature anomaly panel below, minus idio_vol_20, max_return_21d, and beta_60. The prune is justified by the collinearity and ablation analysis that follows (it raises in-regime OOS IC from +0.077 to +0.087, identically for OLS / Lasso / Ridge, which coincide on this panel). The full 9-feature superset (the prior headline lasso_elasso_pit_h21), curated on "one feature per economically distinct anomaly family," with ✓ = kept in the headline 6:

#	Feature	Anomaly family	Reference	In 6?
1	momentum_12_1	Cross-sectional momentum (12-month, skip-1-month)	Jegadeesh-Titman 1993, J. Finance 48(1)	✓
2	momentum_756	Long-horizon trend (36-month)	Moskowitz-Ooi-Pedersen 2012 (TSMOM-adjacent)	✓
3	return_1d	Short-term reversal	Lehmann 1990 / Jegadeesh 1990	✓
4	vol_ewm_20	Low-volatility	Ang-Hodrick-Xing-Zhang 2006	✓
5	idio_vol_20	Idiosyncratic vol (60-day beta residual)	Ang-Hodrick-Xing-Zhang 2006	— dropped
6	max_return_21d	MAX (lottery-stock) effect	Bali-Cakici-Whitelaw 2011	— dropped
7	distance_52w_high	52-week-high anchoring	George-Hwang 2004	✓
8	log_dollar_volume	Size / liquidity	Amihud 2002	✓
9	beta_60	Market exposure control (BAB-adjacent)	Frazzini-Pedersen 2014	— dropped

Net-of-cost decomposition

Gross IC and gross long-short Sharpe assume costless rebalancing. Net-of-cost Sharpe accounts for transaction-cost drag from portfolio turnover. Because the signal targets a 21-day horizon, the book is held over the horizon and per-rebalance turnover is the L1 distance between the long-short weight vector and its value 21 trading days earlier (0.5 · Σ_i |w_t[i] − w_{t-21}[i]|). Annual turnover = mean(per-rebalance turnover) × (252/21) (≈ 12 rebalances/year). After-cost return = gross return − turnover × cost_bps / 10000 × 2, where the × 2 is the round trip (in and out). After-cost Sharpe is computed on the net-return series with the same √(252 / horizon) annualization as gross Sharpe. (Measuring turnover on a 1-day lag — the previous convention — charged daily churn against a 21-day return and over-stated turnover ~21×.)

Reported at 3 bp (institutional all-in), 10 bp (small-fund), and 20 bp (retail-equivalent). Turnover and cost-drag mechanics are implemented in src/price_model/eval/turnover.py. For the regime-confined headline use scripts/net_cost_regime_21d.py — it fits the linear family (OLS / Ridge / Lasso on curated-6) fresh with the temporal IC-scored CV, reads the trees + momentum from the store, and reports the IC point estimate with its HAC-corrected t-stat plus gross/net Sharpe on the non-overlapping 21-day rebalance schedule (the headline convention); scripts/net_cost_regime.py reports the same on the daily-overlap estimator. Avoid the shared scripts/compare_net_of_cost.py: it intersects the full accumulated store (many overlapping historical runs) and drops most of the new regime models.

Audit-driven LightGBM v3 panel

The v3 panel curation (14 features chosen from a 21-feature candidate panel via gain importance + cross-feature correlation analysis) is documented inline in config/experiments/extended_kaggle_v3_curated.yaml; audit script in scripts/audit_lightgbm_features.py.

Discussion

The regime-confined headline finding — a 6-feature linear cross-sectional regression beats every held-out-tuned tree and long-horizon momentum on both gross IC and net-of-cost Sharpe, but only within the post-2022-10 regime it was trained on — raises five questions worth unpacking. This section addresses each.

Why linear beats trees on this universe

The 21-day cross-sectional signal on the post-2024 S&P 500 is approximately linear in feature space. Three pieces of evidence:

1. The momentum factors alone capture most of the signal. mom_504, mom_756, and mom_378 individually produce IC = +0.04 to +0.05 with no model fitting; the OLS regression's incremental edge of +0.02 IC over the best single momentum factor reflects modest additional signal from the non-momentum features it keeps (low-volatility vol_ewm_20, 52-week-high anchoring, liquidity). The dominant component is a linear momentum signal that trees would have to recover via many small splits.
2. Tree-ensemble gain-importance audits show feature concentration. On the 14-feature v3 panel, three features produce 39% of LightGBM gain. The split structure devotes most of its capacity to a handful of features that OLS can express as a small number of coefficients. Trees' theoretical interaction-capturing advantage doesn't pay off here because the dominant signal isn't interactive.
3. Held-out-tuned trees still fall short. With each tree Optuna-tuned (held out at 2024-12-31) on its best panel, the regime-confined OOS IC is +0.069 (LightGBM) / +0.061 (CatBoost) / +0.058 (XGBoost) vs OLS's +0.087 — a ~0.02–0.03 gap.

Two caveats limit how broadly this conclusion generalizes:

• The 6-feature panel is small and ex-ante-curated. The linear advantage depends on this curation. On the broader 14-feature engineered panel the linear recipe is weaker — but so are the trees (every model is worse there), so curation, not model class, is the lever.
• The post-2022-10 regime is structurally momentum-friendly. In a regime with more leadership rotation or weaker momentum persistence, the linear-on-anomalies recipe would compress and trees might catch up. The result is robust within the studied window but has not been verified across regime changes. The conclusion is therefore "within the post-2022-10 regime, on this universe at this horizon, a curated 6-feature OLS outperforms every tree-ensemble variant tested" — not "ML cannot succeed in cross-sectional equity prediction generally," and not "the edge is regime-robust."

Net-of-cost: turnover, and the framing that flips the winner

Net-of-cost ranking depends on turnover, and the two training framings give opposite winners — the gap between them is itself a finding.

Regime-confined (headline): OLS wins net. Trained on the bull regime, the 6-feature OLS model has gross Sharpe +2.09 and, on the 21-day rebalance clock, turns over only ~9×/yr, so at 20 bp round-trip it retains net +1.85. The 36-month momentum factor is the low-turnover champion (4×, net +1.28) but its lower gross (+0.050 IC, Sharpe +1.41) can't catch it. Tuned trees net +0.6–0.8 at 20 bp (turnover 10–11×) — and none is statistically significant once the overlap is HAC-corrected (t < 2). So in-regime, OLS wins on both gross and net — the first configuration in this project where the model beats momentum net-of-cost.

Data quality and methodological limitations

The primary out-of-sample estimate is honest about what it measures, but what it measures is bounded by the data available without a paid feed. This section enumerates those boundaries.

Limitations of yfinance

yfinance is the only free source of daily-bar US equity data and is the reason the project is reproducible without a paid subscription. It has three documented failure modes that materially affect the reported results.

1. Delisted-ticker history is permanently lost. When a company is acquired, fails, or goes private, yfinance stops returning data for the old symbol. SIVB (Silicon Valley Bank, failed March 2023), FRC (First Republic, failed May 2023), SBNY (Signature Bank, failed March 2023), ATVI (Activision, acquired by Microsoft October 2023), AGN (Allergan → AbbVie 2020), and CERN (Cerner → Oracle 2022) have no usable pre-event history accessible via yfinance. Approximately 21 such tickers are enumerated in the project drop list, documented inline in src/price_model/data/tickers.py.
2. Symbol-parser fragility. yfinance fails to fetch some single- and short-letter tickers due to ambiguity with currency or commodity symbols in its routing — for example, K (Kellanova), FI (Fiserv), and DAY (Dayforce), all currently live, exchange-listed US large-caps. yfinance returns "no data" after retries.
3. Foreign listings are unreachable. Acquired companies that consolidated under non-US ADRs cannot be fetched. SIE.DE (Siemens Healthineers, acquired Varian) and MC.PA (LVMH, acquired Tiffany) are examples, despite active trading on European exchanges.

Coverage. Of the 701 ticker symbols that Wikipedia identifies as S&P 500 members at some point during 2017-2026, yfinance provides usable data for 617 (~88%). The remaining ~12% are absent from the panel.

Partial PIT reconstruction from Wikipedia

The Wikipedia "List of S&P 500 companies" page and its "Selected changes" table are the only free source of historical index membership. The scraper in src/price_model/data/sources/sp500_membership.py reconstructs the (ticker, added_date, removed_date) table from those sources. Three known incompleteness modes remain.

1. The change log is reliable only back to approximately 2014. Earlier add / remove events are absent. This does not affect the 2017-start evaluation window directly, but it implies that any pre-2017 extension (for example, applying the same machinery to a 1990s-2010s sample) would inherit incomplete PIT membership.
2. Renames and continuations are ambiguous. When a company changes its ticker (FB → META, RTN → RTX, FISV → FI), Wikipedia logs the event as a simultaneous remove and add on the same date. The scraper treats such pairs as continuous membership at the new symbol; the alias is resolved in tickers.py. The choice is correct for backtest purposes but implies that a researcher reading the membership table directly may miscount index changes by approximately 30 events over the full window.
3. Wikipedia editors can be incorrect, late, or partial. Missing entries (e.g., short-lived 2018 additions never logged as removals) were observed during the build. The membership table is not independently audited against a primary source.

Ticker resolution rules

Three small lookup tables are maintained in src/price_model/data/tickers.py to bridge the gaps above:

Table	Purpose	Size
TICKER_ALIASES	Map renamed symbols to their successor (FB → META, RTN → RTX, ~70 entries documented with rename dates)	73 entries
TICKER_DROP_LIST	Symbols with no usable yfinance data (failed banks, acquired with non-unified history, foreign listings, short-ticker parser failures)	38 entries
SYMBOL_NORMALIZATION	Punctuation convention (BRK.B → BRK-B)	2 entries

The function resolve_ticker(symbol) -> str | None applies all three in order. Every ticker in every universe file is routed through it. Tests in tests/test_tickers.py cover the precedence rules and known edge cases.

These tables are derived heuristically: cli refresh-data is run on the expanded universe and the resulting yfinance failure log is inspected. Each entry carries a one-line comment identifying the corporate event and the rationale for inclusion. New entries are added when new failures are observed.

Sources of upward bias in the reported ICs

The regime-confined headline IC of +0.087 (6-feature OLS on 2025+ OOS) and every other IC in the README — gross or net — should be interpreted with the following caveats. The bias sources are inherent to the data and methodology and apply uniformly across models, not to one model in isolation.

0. Regime-confined training is best-case, by construction. The headline trains only on the bull regime it is then deployed into.
1. PIT correction is partial. Because yfinance does not return data for SIVB, FRC, ATVI, AGN, and similar delisted symbols, those tickers do not appear in the PIT panel even when Wikipedia indicates they were index members on the relevant dates. The PIT-corrected backtest therefore still excludes the worst realized losers of the 2022-2023 banking crisis. A fully PIT-correct analysis using paid data (Norgate, Polygon, or CRSP) would either misrank or correctly short SIVB at −85% on 2023-03-08; the present model does neither.
2. Transaction costs, taxes, and slippage are partially modeled. The 3/10/20 bp net columns model bid-ask + commission only; they exclude slippage on illiquid names, capital-gains taxes, and short-borrow. In-regime the 6-feature OLS nets +1.85 at 20 bp (turnover 9×) and momentum +1.28 (4×) — both institutional-grade gross-and-modeled-net, but the unmodeled frictions and small-account constraints still bite a retail book.
3. Single evaluation window. The 2025+ slice is one window. Alternative windows (2018–2020, 2020–2022, 2022–2024) would produce different ICs; the multi-window stress test is the most-needed extension and has not been performed.

Toward a fully PIT-correct evaluation

Listed in approximate order of effort:

• Replace yfinance with Norgate Premium Data (~$60 / month) for survivorship-bias-free prices including delisted history. All reported ICs would change; the headline edge would likely compress somewhat, and the regime-conditional shape (strong when training matches the deployment regime, fading across regime breaks) is expected to persist.
• Replace the Wikipedia membership source with CRSP (paid; free academic access for most affiliated researchers). Provides cleaner pre-2014 history and eliminates scrape fragility.
• Run multi-window robustness tests at three different start / end pairs.
• Add transaction-cost modeling at the prediction-store layer so reported metrics are net of realistic costs.

Scope and limitations

• Not deployable for retail trading. In-regime the 6-feature OLS model is institutional-grade on both gross (IC +0.087, Sharpe +2.09) and modeled net (Sharpe +1.85 at 20 bp). But the net columns model only bid-ask + commission; once you add slippage on size, capital-gains taxes, short-borrow, and the small portfolio sizes a retail account can hold, the realistic retail edge shrinks sharply. Combined with the regime-bound caveat below, this is a research result, not a trading recommendation.
• The edge is regime-concentrated — and the headline trains inside it. The headline confines training to the post-2022-10 regime by design, so its edge lives entirely in the momentum-persistent regime: on the 2025+ test slice, regime-confined training scores IC +0.087.
• Single evaluation period, no multi-window robustness. All numbers are one OOS window (2025+). A multi-window stress test (2018-2020, 2020-2022, 2022-2024, 2024-2026 in sequence) — ideally with paid PIT-correct data — is the most-needed extension and has not been performed. "Extend the regime" is the natural next step from this branch.
• Survivorship-bias is only partially corrected. The PIT correction excludes ~12% of historical S&P 500 members (delisted symbols yfinance can't fetch). Measured directly (scripts/survivorship_ablation_headline.py), inflation is ≈0 on the 2025+ slice but a sign-flip over the full sample; the absent delisted names make even that a floor. See Data quality and methodological limitations.
• Not a substitute for index funds. For individual investors, decades of research show that low-cost diversified index funds outperform almost all active strategies after fees and taxes.

Architecture

src/price_model/
├── data/
│   ├── sources/         # yfinance + Ken French + Wikipedia adapters
│   ├── universes/       # static universe files (sp500.txt, sp500_pit.txt)
│   ├── tickers.py       # aliases / drop list / normalization rules
│   ├── membership.py    # PIT membership lookup + filter
│   └── loaders.py       # one-call load_panel(universe, start, pit_filter)
├── features/            # technical, cross-sectional, factor-loading, anomalies
├── models/              # LightGBM, baselines, classical (ARIMA, GARCH, GBM, FF)
├── pipeline/            # walk-forward backtest harness
├── eval/                # IC, hit rate, Sharpe, bootstrap CI, time-split
├── serving/             # DuckDB prediction store
├── dashboard/           # Streamlit dashboard reading from the store
└── cli.py               # `price-model` entry point

config/experiments/      # YAML configs for each stage above
notebooks/               # diagnostic + classical + robustness + portfolio + features
tests/                   # leakage tests, PIT tests, ticker tests, contract tests

Notebooks (each maps to one or more README sections; together they provide interactive support for the project's claims):

• notebooks/00_pit_ablation_study.ipynb — PIT ablation study on the legacy v2 LightGBM panel. Quantifies survivorship-bias inflation (61–89% depending on feature-set strength). Supports Methodology + Data quality.
• notebooks/01_prediction_diagnostics.ipynb — Stored-prediction diagnostics on the headline 21-day models: apples-to-apples comparison, rolling 60-day IC, regime-conditional IC, net-of-cost gross-vs-net Sharpe with turnover annotation. Supports the net-of-cost and regime-conditionality analysis in Discussion.
• notebooks/02_classical_timeseries.ipynb — Classical EMH baselines (ARIMA / GARCH / GBM) on the per-ticker time series, contrasted with the cross-sectional linear model. Supports the case that the cross-sectional approach beats per-ticker time-series baselines.
• notebooks/03_headline_robustness.ipynb — Bootstrap CI / decile bucket monotonicity / time-split partition on the linear headline. Supports the regime-conditionality analysis.
• notebooks/05_feature_exploration.ipynb — Feature engineering boilerplate: 41-feature registry tour, distributions, correlations, fitted L1 / LightGBM importance, and a step-by-step template for adding new features. Now also includes the panel-ablation findings (9→6 prune, beta_60 anti-generalizing, regularization-inert OLS≡Ridge≡Lasso, vol-cluster-is-signal). Supports Methodology (panel design) and the panel-ablation findings (9→6 prune; regularization-inert OLS ≡ Ridge ≡ Lasso).
• notebooks/06_hp_selection_bias.ipynb — Visualizes the cross-regime HP fragility: chosen HPs swinging across training cutoffs (lambda_l1 over ~two orders of magnitude) and the held-out-vs-default inversion at 21-day. Contrast with the in-regime sweeps in Methodology's Optuna protocol, where CV predicts OOS.

Extras (independent of the cross-sectional ML narrative):

• notebooks/07_portfolio_attribution.ipynb — Fama-French 5-factor risk decomposition + realized return attribution on a fixed portfolio. Reuses the project's FF5 data adapter and rolling-beta features for an orthogonal use case (portfolio risk, not return prediction). Independent application; the numbering jumps to 07 to signal scope drift from notebooks 00-06.

Streamlit dashboard

src/price_model/dashboard/ is a thin Streamlit reader over the DuckDB prediction store. It exists to allow the model's daily output to be inspected visually (per-date top / bottom quintile, rolling IC, prediction vs. realized scatter) without writing a notebook for each inspection. The dashboard is a debugging and monitoring surface, not part of the reproduction workflow. None of the reported numbers in this README originate from the dashboard, and skipping it does not affect reproduction of any ablation cell. The dashboard is most useful when extending the model and a quick sanity-check view of new runs is desired.

Development notes

The substantive research decisions in this project — what to measure, what constitutes a fair test, how much survivorship bias the primary estimate contains, which limitations are honest to ship with — are the author's. The engineering side of the project (test scaffolding, CI configuration, package layout, refactors, and the translation from research intent to reproducible implementation) was developed in close collaboration with Claude Code. Specifically:

• The walk-forward harness, prediction store, and dashboard scaffold were designed and iterated in conversation with Claude Code.
• Test coverage (leakage tests, PIT-membership tests, contract tests on the loaders, ticker-resolver tests) was co-designed; Claude Code drafted many of the test cases that subsequently surfaced bugs in the implementation.
• The GitHub Actions CI pipeline (ruff + pyright + pytest on Python 3.11) and the pre-commit hooks were configured with Claude Code.
• Code review, refactors, and a final pre-publish audit pass were conducted in collaboration with Claude Code.

The README, the experiment configs, and the ablation narrative are the author's claims; the engineering quality of the surrounding codebase is a product of the above collaboration.

Citations

• Han, Y., He, A., Rapach, D., and Zhou, G. (2024). "Expected Stock Returns and the Cross-Section: An E-LASSO Approach." Review of Finance. — Panel-curation principle (one feature per economic family, rank-normalized, CV-selected α) used in the headline 6-feature linear regression (lasso_curated6_pit_h21), pruned from a 9-feature superset.
• Jegadeesh, N. and Titman, S. (1993). "Returns to Buying Winners and Selling Losers." Journal of Finance 48(1). — 12-1 momentum anomaly.
• Moskowitz, T., Ooi, Y. H., and Pedersen, L. H. (2012). "Time Series Momentum." Journal of Financial Economics 104(2). — Long-horizon (TSMOM) momentum.
• Lehmann, B. (1990). "Fads, Martingales, and Market Efficiency." Quarterly Journal of Economics 105(1). — 1-day reversal.
• Ang, A., Hodrick, R., Xing, Y., and Zhang, X. (2006). "The Cross-Section of Volatility and Expected Returns." Journal of Finance 61(1). — Low-vol and idiosyncratic-vol anomalies.
• Bali, T., Cakici, N., and Whitelaw, R. (2011). "Maxing Out: Stocks as Lotteries and the Cross-Section of Expected Returns." Journal of Financial Economics 99(2). — MAX (lottery) effect.
• George, T. and Hwang, C.-Y. (2004). "The 52-Week High and Momentum Investing." Journal of Finance 59(5). — 52-week-high anchoring.
• Amihud, Y. (2002). "Illiquidity and Stock Returns: Cross-Section and Time-Series Effects." Journal of Financial Markets 5(1). — Liquidity / size proxy.
• Frazzini, A. and Pedersen, L. H. (2014). "Betting Against Beta." Journal of Financial Economics 111(1). — Market-beta control.
• Fama, E. and French, K. (2015). "A Five-Factor Asset Pricing Model." Journal of Financial Economics 116(1). — FF5 baseline (reference only; not in the headline comparison due to Kenneth French release-lag truncation).
• Zou, H. and Hastie, T. (2005). "Regularization and Variable Selection via the Elastic Net." Journal of the Royal Statistical Society, Series B 67(2). — ElasticNet model class.
• Bailey, D. and López de Prado, M. (2014). "The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality." Journal of Portfolio Management 40(5). — Multi-test Sharpe correction.
• DeMiguel, V., Garlappi, L., and Uppal, R. (2009). "Optimal Versus Naive Diversification: How Inefficient is the 1/N Portfolio Strategy?" Review of Financial Studies 22(5). — The estimation-error-dominates-optimization result underlying the cross-regime HP-fragility caveat.
• López de Prado, M. (2018). Advances in Financial Machine Learning. Wiley. — Purged walk-forward cross-validation (Ch. 7).
• Ken French Data Library. Daily factor returns. https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html
• Wikipedia: List of S&P 500 companies. Historical components and change log. https://en.wikipedia.org/wiki/List_of_S%26P_500_companies

License

MIT.