Broker Data Differences and the Reproducibility of Trading Models

Summary

The article argues that decentralized broker feeds can produce materially different histories for the same forex symbol, making trading research and machine learning results difficult to reproduce across brokers. It compares four years of daily EURUSD data from two brokers, examining timestamp alignment, returns, variance, correlations, and forecast error relative to a mean-return baseline. The reported comparisons include a correlation of 0.41 and differences in average returns, variance, forecastability, and tail behavior.

The author uses these discrepancies to argue that models trained or optimized on one broker’s data may perform differently when deployed through another, and suggests broker-specific development and evaluation. The discussion also calls attention to tick volume and trading-session conventions as sources of variation. Its evidence comes from only two anonymized brokers, one currency pair, and a particular historical sample. Misaligned timestamps and other data-handling choices may affect the comparisons, so the results motivate careful validation rather than establishing that broker-specific models are always necessary.

Key ideas

• Broker feeds may differ in timestamps, prices, spreads, and tick volumes even for the same symbol and requested history length.
• The article compares broker EURUSD returns using distribution statistics, correlation, and forecast error against a mean-return baseline.
• Its sample reports a broker-to-broker return correlation of 0.41 and differences in several statistical properties.
• Models trained on one broker’s data may not reproduce the same results on another broker’s feed.
• The two-broker, single-pair comparison is limited and sensitive to timestamp alignment and other data-processing choices.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.