Skip to content

preregistration.md

Latest commit

 

History

History
132 lines (107 loc) · 6.73 KB

preregistration.md

File metadata and controls

132 lines (107 loc) · 6.73 KB
Raw

Prospective Evaluation Protocol

Status: repository-frozen prospective protocol. Commit 4018f4063f46889f41d6981db5a71079e1dbd713 and protocol SHA-256 e49e41a12a19fa5404a573ba5e21eb8a2888e616985f8c610d9652866923315c are the pre-retrieval public record. This is not an external registry entry. The 2018-2025 six-ETF case in the paper was inspected before this freeze and is not confirmatory evidence for the expansion.

Research Question

How much do explicit timing, cash, cost, comparator, and engine contracts change reported performance across heterogeneous target-weight strategies and real data panels?

Frozen Panels

Both panels use daily adjusted closes from 2014-01-02 through 2025-12-31 and SHV as the residual-cash proxy. Evaluation begins 2018-01-02 and ends 2025-12-31. Raw matrices remain untracked; each run records provider metadata, retrieval time, input SHA-256, and row coverage.

  1. us_sector_etfs: XLB, XLE, XLF, XLI, XLK, XLP, XLU, XLV, XLY.
  2. country_equity_etfs: EWA, EWC, EWG, EWH, EWJ, EWL, EWP, EWQ, EWS, EWU.

The accepted matrix is the complete-row intersection of the declared symbols and SHV. There is no forward fill or symbol substitution. A panel is excluded only if it has fewer than 252 complete pre-evaluation sessions or a missing evaluation month. The learned model additionally requires 24 mature monthly training dates before evaluation. Any exclusion is reported as a protocol deviation; a new universe is not substituted after outcomes are inspected. Provider terms and publication rights are reported per panel without legal inference.

The frozen retrieval adapter is yfinance with auto_adjust=False, actions=False, repair=False, and threads=False; the adjusted-close field is selected explicitly. The request ends at 2026-01-01 because the provider's end date is exclusive. This choice does not assert permission to redistribute the observations.

Frozen Strategies

All decisions occur on the first observed session of each calendar month after every feature is mature. Signals use that session's close and execute on the first strictly later panel row.

  • ts_momentum: 252-session total return. Each positive-signal asset receives 1 / N_positive; otherwise capital remains in SHV.
  • cross_sectional_momentum: return from session t-252 through t-21. The top three assets receive one-third each, irrespective of sign.
  • short_term_reversal: negative five-session return. Up to the three assets with the lowest negative returns receive one-third each; unused exposure remains in SHV.
  • learned_gbrt: a walk-forward gradient-boosted regression model predicts 21-session forward log return. Features are 5-, 21-, 63-, 126-, and 252-session log returns; sample standard deviations of daily log returns over 21 and 63 sessions; price / trailing_252_session_high - 1; and normalized ordinal cross-sectional ranks of 21- and 252-session returns. The lowest rank is zero and the highest is one. Training examples are prior monthly decisions whose labels end on or before the current decision. The pooled training rows are asset-month pairs; symbol identity is not a feature. The rolling training set is capped at 60 decision months and must contain at least 24. The estimator is scikit-learn GradientBoostingRegressor with 100 estimators, learning rate 0.03, depth 2, minimum leaf size 10, and subsample 0.8. Seeds are 11, 29, 47, 71, and 97. Each seed is reported; the family estimate is the arithmetic mean of seed-level metrics. Up to three positive predictions receive one-third each.

All score and rank ties are broken by ascending symbol. Features are not standardized or winsorized. No hyperparameter, universe, threshold, seed, or window may change in response to observed performance.

Execution And Comparators

Primary accounting uses the event-driven engine, initial NAV 1.0, next-bar close execution, SHV residual-cash returns, and 13 basis points per unit of one-way gross traded notional. Targets are long-only with gross exposure at most one. The vectorized engine is a model-convention sensitivity diagnostic on the identical canonical target tape; equality with pseudo-share accounting is not expected.

For each strategy, a causal costed comparator follows the strategy's target risky exposure, allocates it equally across the panel, rebalances on the same dates, holds residual SHV, and pays the same cost rate. A buy-and-hold equal weight panel portfolio is descriptive only.

The one-switch diagnostics are:

  1. next-bar close versus deliberately invalid same-close assignment;
  2. SHV residual cash versus zero cash return;
  3. 13 basis-point cost versus zero cost;
  4. raw net return versus the causal exposure-matched comparator; and
  5. event-driven versus vectorized accounting.

The invalid same-close result is a diagnostic, not a candidate strategy.

Outcomes And Uncertainty

Primary outcomes are the paired change in annualized arithmetic return and the paired change in conventional sample Sharpe for each switch, strategy family, and panel. Annualized arithmetic return is 252 times the daily mean. Sharpe is the mean daily strategy return minus the SHV return, divided by its sample standard deviation and multiplied by the square root of 252. Secondary outcomes are net CAGR, maximum drawdown, turnover, cost drag, active return versus the costed comparator, engine return divergence, strategy-rank reversals, and learned-model seed dispersion.

Joint circular-block bootstrap intervals use 5,000 draws, primary block length 21 sessions, sensitivity lengths 5 and 63, and seed 20260618. Return components are resampled jointly. Exact accounting identities must hold in every original series and sampled draw. Intervals are percentile intervals and are not called posterior probabilities or multiple-testing-adjusted evidence.

Researcher Degrees Of Freedom

  1. Archive all configurations, seeds, failures, exclusions, and deviations.
  2. Preserve the current negative case without retuning it.
  3. Label analyses outside this protocol exploratory.
  4. Do not call any period inspected before this commit out of sample.
  5. Publish the machine-readable protocol, canonical target-tape hashes, aggregate result tables, forest plots, engine-conformance matrix, rank-reversal summary, environment lock, and content hashes.
  6. Do not publish raw provider matrices unless explicit redistribution rights are documented.

The machine-readable source of truth is paper/expansion/protocol.json.

Post-freeze clarification: the JSON key cost_per_one_way_gross_notional is a naming error retained to preserve the pre-retrieval protocol hash. The implementation applies its 0.0013 value to gross two-sided traded notional, sum(abs(delta_weight)), not to one-way turnover. This clarification changes no calculation or frozen parameter.