rodynaemad/btc-random-walk-benchmark

Bitcoin vs. the Random Walk: Out-of-Sample Forecasts of 33 Models

Every out-of-sample prediction behind the working paper "Can Machine Learning Beat a Random Walk in Bitcoin? A Leakage-Free Benchmark of 32 Models" (Rodayna Emad Mohamed, 2026), plus the result tables computed from them.

Use it to reproduce the paper's tables, run your own significance tests or backtests, or test whether a new model clears the same bar without retraining the 33 models.

• Code and paper: https://github.com/itsRou/bitcoin-price-prediction
• Interactive explorer: https://huggingface.co/spaces/rodynaemad/bitcoin-random-walk-benchmark

Headline result

After Holm correction for 32 comparisons, no model is significantly more accurate than the zero-return (random-walk) forecast at either horizon (Diebold–Mariano test), while 23 of 32 (1-day) and 25 of 32 (7-day) are significantly worse. Directional accuracy peaks at 51.1%, and the one trading strategy that beats buy-and-hold after costs does not survive a deflated Sharpe ratio test.

Contents

File	Rows	Description
predictions/oof_h1.csv	3,359	1-day forecasts of all 33 models
predictions/oof_h7.csv	3,353	7-day forecasts of all 33 models
predictions/oof_h1_stationary_features.csv	3,359	robustness run: 16 models (plus the zero forecast) without price-level features
results/table_h1.csv, results/table_h7.csv	33	per-model RMSE, out-of-sample R², directional accuracy, DM tests, Holm-adjusted p-values
results/backtest_h1.csv	61	cost-aware long/short and long/flat backtests
results/fold_da_h1.csv	13	directional accuracy per walk-forward fold
results/summary*.json	—	headline numbers quoted in the paper

Prediction file columns

• timestamp — UTC date t (the forecast is made with information up to the close of t − 1)
• y_true — realized forward log return, ln(P_t+h) − ln(P_t)
• fold — walk-forward test fold (0–12). Rows with an empty fold were never scored (initial training period) and should be dropped for evaluation.
• one column per model with its out-of-sample forecast of y_true

Models: zero return, mean return, always long, lag-1 OLS, auto-ARIMA, Holt–Winters, Prophet, OLS, ridge, lasso, elastic net, Bayesian ridge, Huber, k-NN, SVR, decision tree, random forest, extra trees, AdaBoost, gradient boosting, XGBoost, LightGBM, CatBoost, MLP, RNN, LSTM, GRU, BiLSTM, 1-D CNN, CNN–LSTM, LSTM with attention, Transformer and TCN.

How the predictions were produced

• Data: daily BTC-USD candles, 1 Jan 2017 – 30 Sep 2026 (Yahoo Finance).
• Features: 66 OHLCV-derived features (lagged returns, rolling statistics, volatility estimators, technical indicators, regime descriptors), all shifted one day to prevent look-ahead.
• Validation: expanding-window walk-forward CV with purging and embargo (17 blocks); only the 13 folds with ≥ 2 years of training history are scored (20 Sep 2019 – 29 Sep 2026).
• Models: library defaults, no tuning on test periods; deep models trained with early stopping.

import pandas as pd
df = pd.read_csv("hf://datasets/rodynaemad/btc-random-walk-benchmark/predictions/oof_h1.csv",
                 index_col=0, parse_dates=True)
scored = df[df["fold"].notna()]
r2_os = 1 - ((scored.y_true - scored.xgboost) ** 2).sum() / (scored.y_true ** 2).sum()
print(f"XGBoost out-of-sample R² vs random walk: {r2_os:.3%}")

Notes and limitations

• Raw OHLCV prices are not redistributed here; the repository fetches them from Yahoo Finance and freezes a snapshot for exact reproduction.
• Results cover one asset, one daily frequency and one seven-year history, with OHLCV features only. See the paper's limitations section.
• Research and education only; not investment advice.

Citation

@misc{mohamed2026bitcoinrandomwalk,
  author = {Rodayna Emad Mohamed},
  title  = {Can Machine Learning Beat a Random Walk in Bitcoin? A Leakage-Free Benchmark of 32 Models},
  year   = {2026},
  note   = {Working paper},
  url    = {https://github.com/itsRou/bitcoin-price-prediction}
}