Statistically Detectable, Economically Small: A Pre-Registered Out-of-Sample Test of Minute-Level Bitcoin Return Predictability
Year: 2026
Type: preprint
Abstract: We test whether a fixed forecasting program built from public one-minute candles of Binance BTCUSDT spot carries out-of-sample information about 15-minute returns, and how much that information is worth. The program (fourteen causal features, Ridge regression, expanding window) and its tests were frozen and hashed before the evaluation data were read. A historical external validation over 56 monthly folds (2021-2025, 2.45 million minutes) rejects the null of no predictive information (t=5.44), but the evidence is uneven: 76% of the summed Clark-West differential comes from 2021 (excluding 2021, t=3.00), and a test on the 56 monthly means gives t=2.14. A within-month demeaned test (t=5.74) shows that the forecasts carry timing information rather than agreement with the sample trend. Two confirmatory holdouts, each run under a protocol frozen before its data were opened, split. July-August 2026 rejects (t=1.89, p=0.029). September-October 2026, with a drift-free primary statistic, does not (t=1.33, p=0.092), although its estimate (an information coefficient of 0.021) is close to the historical one (0.024). Economically the information is small: squared-error tests never reject, the 95% upper bound on the out-of-sample R^2 is about 0.15-0.2%, and a non-overlapping 15-minute rule breaks even at 0.2-0.3 basis points per side, against a retail taker fee of 10 basis points. A census of 2,070 days of Binance archives finds 157 minutes where the exchange's monthly and daily files disagree; excluding them leaves the results unchanged. Short-horizon predictability of BTCUSDT from public candles is weak, fragile in short samples and economically small.
Author: Zeng, Gaoshan
Language: English
Subfield: Finance
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7236132351
