Cryptocurrency Exchange Rate Forecasting: Robustness of Binary Classifiers

A new study examines the robustness of binary classifiers in cryptocurrency exchange rate forecasts. The researchers used three price-ratio series formed from MLNUSDT, PLUMEUSDT, and SIRENUSDT perpetual-futures trades on Binance Data Vision. They combined a Light Gradient Boosting Machine (LightGBM) classifier with two transparent conditions: a positive trend-continuation cue and a rolling Hurst-regime threshold.

The study found that event-time relative-value ratios provided repeatable forward classification behavior in the short sample. However, the researchers cautioned that apparent improvements may not survive component ablation, alternative classifiers, multiple seeds, walk-forward evaluation, and trading-cost stress tests.

The study used a revised design to address repeated specification search, dependent observations, and execution costs. It reported precision, recall, F1, average precision, balanced accuracy, positive-signal coverage, uncertainty, and simple cost-adjusted accepted-signal returns. The researchers also compared standardized logistic regression, random forest, XGBoost, and LightGBM under a common feature set and chronological evaluation.