Meta Description: A practical anti-overfitting checklist for trading strategies: trial accounting, deflated Sharpe, purged CV, walk-forward, and PIT discipline — with Python code for each gate. The five-gate framework from López de Prado (2018).

Most backtested strategies fail in production for one reason: overfitting — the research process memorized history instead of discovering structure. Avoiding it is not a single test but a five-gate pipeline, where each gate catches a different failure mode and a strategy must pass all five before paper trading.

For the deep treatments see deflated Sharpe, CPCV, purged CV, walk-forward, and the survivorship audit. This article is the operative checklist that binds them.

1. Count every trial before you compute anything

Overfitting begins the moment the second configuration is tested. Every parameter value, indicator variant, universe tweak, and restarted experiment is a trial, and the reported Sharpe is the maximum over trials — a statistic whose null distribution shifts upward with trial count. The working rule from Harvey, Liu & Zhu (2016): with 100+ trials, a t-ratio below ~3.0 is indistinguishable from luck, not the textbook 2.0.

import numpy as np
from scipy import stats

def trials_needed_for_luck(target_sharpe: float, n_obs: int = 1260) -> int:
    """Trials N at which E[max Sharpe | null] reaches target_sharpe."""
    se = 1.0 / np.sqrt(n_obs)
    for n in (1, 10, 50, 100, 500, 1000, 5000):
        e_max = (1 - np.euler_gamma) * stats.norm.ppf(1 - 1 / max(n, 2))
        e_max += np.euler_gamma * stats.norm.ppf(1 - 1 / (max(n, 2) * np.e))
        if e_max * se * np.sqrt(n_obs) / 1.0 >= target_sharpe:
            return n
    return 5000

Practical compliance is process, not willpower: log every experiment with seed, config, and timestamp in an immutable tracker; freeze the grid before validation; report N alongside every Sharpe. A strategy with undisclosed N has no significance claim at all.

2. Deflate the Sharpe for the trials you ran

Given disclosed N, the deflated Sharpe ratio (Bailey & López de Prado, 2014) computes the p-value of the observed Sharpe against the expected maximum under the null — corrected for skew and kurtosis, which punish crash-risk profiles. A headline Sharpe 1.0 over five years looks significant naively (p ≈ 0.01) yet routinely fails deflation at N=100. Run the deflation before celebrating; most celebrations end here, which is the point.

3. Purge the cross-validation or it validates leakage

Random k-fold on financial data is a leakage machine: overlapping labels and autocorrelated rows let the model train on its test set's neighbours. Purged k-fold with embargo removes training rows whose label windows intersect test blocks plus a serial-dependence buffer. The random-minus-purged score gap is your leakage estimate — if purging halves the Sharpe, half the strategy was the validator, not the market.

Two further rules: all feature selection and standardization happen inside folds (never on the full sample), and hyperparameter search itself runs under purged CV, or the search optimizes for leak exploitation.

4. Walk the strategy forward before trusting any number

Purged CV validates the model; walk-forward validates the production process — re-fit on history, trade frozen configs on unseen segments, stitch only out-of-sample segments into the equity curve. Demand 5–10 OOS steps across regimes, costs inside the selection criterion, and per-window diagnostics (config stability, decay trajectory). A walk-forward Sharpe 30–60% below the best in-sample is normal; near zero means the edge was search artefact.

5. Putting the gates in order: a worked triage

Take a concrete candidate: a momentum variant with headline Sharpe 1.3 over six years of daily data, found after testing roughly 150 configurations. Gate zero (data): the universe uses current index constituents — fail. Rebuild survivorship-free; Sharpe drops to 1.0. That single correction erased a quarter of the edge before any statistics, which is typical: data leakage is the largest and least discussed component of most headline Sharpes.

Gate one (trials): N=150 disclosed. Gate two (deflate): DSR with skew −0.4, kurtosis 5 → deflated p ≈ 0.09 — not significant at 5%. The triage could stop here: the strategy, as specified, does not clear honest significance. But suppose the researcher restricts the claim to a sub-period with stronger economics and N=20 (documented subset, not cherry-pick): deflated p ≈ 0.03 — passes. Gate three (purge): random k-fold Sharpe 0.9 vs purged 0.55 — leakage present but not fatal. Gate four (walk-forward): stitched OOS 0.4 across eight windows with stable config family — modest, real, deployable at reduced size with a regime filter.

Note what happened: each gate shrank the claim (1.3 → 1.0 → significant-only-in-subset → 0.55 purged → 0.4 walk-forward), and the final claim is roughly one-third of the headline. That haircut is the normal fate of searched strategies, and a pipeline that applies it mechanically is worth more than any single model improvement. Strategies that survive all five gates are rare; size them accordingly when you find one.

6. Clean the data last in order but first in importance

None of the above matters on contaminated inputs: survivorship-biased universes, unadjusted corporate actions, unlagged fundamentals, and revised (not vintage) macro all inflate the Sharpe before statistics ever see it. The survivorship and look-ahead audit enforces the three-leg point-in-time discipline — universe membership, prices, and features each knowable at decision time — with mechanical checks (timestamp proofs, delisting sensitivity, ALFRED vintages). Clean data is gate zero: run it before gate one, because a significant Sharpe on leaked data is precisely measured fiction.