Transaction Costs Inflate False Positives in a Day-Shuffle Monte Carlo Test of Trading Signals

AIWMC Research Foundation — AIWMC Research Foundation · AIWMC LLC

AIWMC-RF-TN-2026-001 v1.0 · November 2026 · CC BY 4.0

Technical note. AIWMC's return-predictor research code judges a trading signal by comparing its out-of-sample Sharpe ratio with 400 copies in which the signal's days are randomly shuffled. On 200 simulated random walks, where no strategy can have an edge, the test reported significance at the 5% level in 8.0% (logistic model) and 10.5% (random forest) of runs, and at the 10% level in 16.0% and 19.0%; the 95% intervals at the 10% level exclude the nominal rate. Setting transaction costs to zero restored nominal rates (5.0% and 4.0% at the 5% level), isolating the cause: shuffling breaks up runs of consecutive positions, so shuffled signals trade more and pay more fees, which makes the real signal look better than chance. A circular-shift null that keeps the signal's run structure held nominal rates (5.0% and 3.0% at the 5% level) and still detected weak real predictability in AR(1) return series. Synthetic data only; one cost level, threshold and series length. Scripts and raw outputs are in the research repository.