A Frequency-Controlled Comparison of Tick- and Minute-Based Information Bars for Cryptocurrency Markets

Battle Count: 63

Highly relevant to quantitative trading. The paper directly addresses a critical practical decision: whether to construct information bars from tick or minute data for ML-based trading strategies. Key findings with direct trading implications: (1) tick Renko bars achieve near-perfect random-walk conformance (|VR(4)-1|=0.020, |AC1|=0.002), ideal for reducing serial correlation bias in model training; (2) tick volatility bars reduce serial dependence by 69% vs. minute baseline; (3) tick hybrid bars achieve lowest tail risk among tick pipelines (kurtosis 6.77); (4) matched-frequency analysis confirms tick advantage is real but masked by sampling-frequency artefacts; (5) bar-type-specific guidance is provided for ML practitioners. However, the downstream ML experiment shows near-chance AUC for all bar types, cautioning that statistical bar quality does not guarantee directional predictability. The paper is primarily about data representation quality rather than alpha generation.

Moderate-to-high complexity. Requires: (1) processing tens of millions of tick records (Binance aggTrade) with sequential streaming to maintain accumulators across batches; (2) implementing six distinct bar construction algorithms with different accumulation logic for both tick and minute pipelines; (3) adaptive EMA threshold calibration with self-monitoring stability routines and bounded updates; (4) duration constraint enforcement with timeout tracking; (5) eight statistical evaluation criteria plus supplementary diagnostics (BDS, ARCH-LM, KS, Mann-Whitney); (6) matched-frequency coarsening with phase-averaging; (7) walk-forward CV with embargo gaps for the ML experiment; (8) 28-feature engineering including fractional differencing. The tick pipeline is computationally demanding due to per-trade processing. However, the core bar construction logic is well-specified with explicit equations, and scikit-learn classifiers use default parameters.

The paper states that the full implementation is released as replication materials. Data is sourced from public Binance API (aggTrade and 1-minute OHLCV endpoints) for BTCUSDT USDT-margined perpetual futures. All statistical tests, calibration procedures, and evaluation criteria are fully specified with equations. Walk-forward CV protocol, feature set (28 features), classifier configurations (scikit-learn defaults, random_state=42), and transaction cost assumptions are explicitly documented. The matched-frequency robustness procedure is described in detail. Minor gap: no explicit GitHub URL is provided in the extract, though replication materials are referenced.