An Empirical Evaluation of Artificial Intelligence and Machine Learning–Based Trading Strategies in Cryptocurrency Markets

Abstract

This DBA thesis investigates whether artificial intelligence (AI) and machine learning (ML) based trading strategies provide statistically significant and economically meaningful benefits over traditional rule-based benchmarks in cryptocurrency markets. The study tests three hypotheses across four Binance perpetual futures contracts, BTCUSDT, ETHUSDT, SOLUSDT and XRPUSDT, using hourly data over a common walk-forward out-of-sample period from 2 April 2022 to 1 January 2026, with 15 non-overlapping folds per asset.
The empirical analysis compares AI/ML strategies with rule-based benchmarks on risk-adjusted return, evaluates whether a regime-aware LightGBM strategy reduces maximum drawdown relative to static strategies, and examines whether ensemble-based AI strategies improve Sharpe stability and cross-asset consistency over single-model approaches. The methodology uses train-validation-test walk-forward folds with embargo controls, train-only feature scaling, transaction-cost-aware backtesting, bootstrap confidence intervals, Wilcoxon and Welch tests, Monte Carlo block bootstrap, cost-stress testing, leave-one-asset-out analysis, regime breakdown, and independent audit review.
The results do not support any of the three hypotheses on their primary criteria. AI/ML strategies do not achieve statistically significant higher mean Sharpe ratios than rule-based benchmarks. The pooled rule-based advantage is mainly driven by two exceptional bullish outlier folds; after removing them, performance is nearly identical. The regime-aware strategy underperforms the static benchmark, with mean maximum drawdown of 67.25% versus 37.58%. Ensemble strategies also fail the primary Sharpe-stability test, with the pooled median Sharpe delta favouring single-model strategies.
The findings are not entirely negative. AI/ML strategies record higher median Sharpe ratios across all four assets and perform consistently better in bearish regimes. Aggregation of ensembles reduces maximum drawdown for all four assets, indicating that these strategies may play an important role in preserving the capital when the Sharpe is not improved. In summary, AI/ML strategies do not show universal superiority within the tested four-asset, backtest-only crypto universe, but may add value under specific regimes and architectures. The thesis contributes an empirical, auditable framework for testing AI/ML trading claims under realistic and cost-aware conditions.