Reading Strategy Correlation Heatmaps Across Backtest Trials

Summary

The document explains how to interpret a correlation heatmap associated with many strategy backtests in a financial research paper. The questioner is unsure how thousands of index-versus-strategy data points are ordered in the exhibit. The accepted explanation corrects that interpretation: each row and column represents one strategy trial, and the heatmap entries show correlations between the return series of different trials. Thus the displayed matrix is not a set of index-versus-strategy scatter points.

The answer also notes that the paper’s cluster analysis on the following page builds on this matrix, which makes the trial-to-trial correlations useful for examining similarity among strategies. The document gives no details of the heatmap’s ordering, color scale, correlation calculation, or clustering procedure, so it resolves what the axes represent rather than fully explaining how to reproduce or analyze the exhibit.

Key ideas

• Each heatmap row and column corresponds to a strategy trial.
• The cells encode correlations between returns from pairs of strategy trials.
• The matrix does not depict index-versus-strategy scatter data.
• The subsequent cluster analysis uses the strategy correlation structure.

Tags

# Comprehending Corr heatmap from multiple trials of investment strategy (Paper)


# Comprehending Corr heatmap from multiple trials of investment strategy (Paper)












Can someone please explain the heatmap of Corr from backtests in this research paper ? A Data Science Solution to the Multiple-Testing Crisis in Financial Research - Lopez De Prado (2019) - (Exhibit #3 on Page 3)

It is not clear how the Index vs Strategy correlation data points from 6,385 trials are represented/ordered in the Heat map.

The following explanation in the paper did not help much : "As is customary in machine learning applications, the main diagonal crosses the Cartesian product from the bottom left to the top right."

## Answer by Quantoisseur (score 1, accepted)

https://quant.stackexchange.com/a/57317

Each row/column of the heatmap is a trial. The matrix is the correlations between the different strategy trial returns, not the index vs strategy scatter plot.

The cluster analysis on the next page builds upon this.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

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