The Toolkit

200 strategies with no edge at all, tested on two years of daily data. The best one on the first year has a Sharpe ratio of 1.74; on the second year the same strategy scores −1.07, and the average out of sample is −0.07. Selecting on in-sample results is selecting on luck; only the out-of-sample period says anything.

• Re-optimising at every walk-forward step and then reporting the combined result is still a search; count every variant you tried.
• Short out-of-sample windows are noisy; one good year proves little.
• Walk-forward cannot fix data problems such as survivorship or look-ahead in the inputs.

rng = np.random.default_rng(11)
r = rng.normal(0, 0.01, (200, 1_000))               # 200 strategies with no edge, 1,000 days
ins, outs = r[:, :500], r[:, 500:]
sharpe = lambda x: x.mean(axis=-1) / x.std(axis=-1) * np.sqrt(252)
best = np.argmax(sharpe(ins))
result = (round(sharpe(ins)[best], 2), round(sharpe(outs)[best], 2), round(sharpe(outs).mean(), 2))

Assumes import numpy as np and import pandas as pd (and SciPy where imported). Each snippet simulates or makes up its own data, as the comments say.

# Python — walk-forward with TimeSeriesSplit
from sklearn.model_selection import TimeSeriesSplit

tscv = TimeSeriesSplit(n_splits=5)
results = []
for train_idx, test_idx in tscv.split(X):
    X_train, X_test = X[train_idx], X[test_idx]
    model.fit(X_train, y[train_idx])
    score = model.score(X_test, y[test_idx])
    results.append(score)

print(f"Walk-forward: {np.mean(results):.3f}")