Data preprocessing

López de Prado's methods, step by step

2026MIAX master's, Instituto BME

Figure 1. The same close series as time, tick, volume and dollar bars

Summary

Tick, volume and dollar bars, fractional differentiation, covariance cleaning with Marchenko–Pastur, triple-barrier labels and splits that respect time. Dollar bars were chosen as the base, and the series turns stationary from d = 0.4 while keeping part of its memory. Done in a pair.

Bars and memory

López de Prado's preprocessing methods, step by step. Tick, volume and dollar bars are built and compared, and dollar bars become the base. Fractional differentiation is tried from d = 0 to 1; from d = 0.4 the series is stationary while keeping part of its memory.

Covariance, labels and splits

A covariance matrix of ten cryptocurrencies is cleaned with Marchenko–Pastur, events are labelled with the triple barrier, using fixed and volatility-based thresholds, and the train and test splits respect time.

Figures

**Figure 2.** Fractional differentiation by degree d: stationary from d = 0.4

**Figure 3.** Triple-barrier labels for three thresholds

**Figure 4.** Purged cross-validation with an embargo, fold by fold

Built with

PythonpandasNumPyscikit-learnParquetJupyter