Correlation-Based Pattern Matching for Online Portfolio Selection
Summary
The document explains online portfolio strategies that find historical market windows resembling current conditions. CORN measures similarity with Pearson correlation rather than Euclidean distance and uses the resulting matches to guide portfolio weights. It also describes ensembles: CORN-U spreads capital evenly across experts with different parameters, while CORN-K allocates evenly among the best recent performers. SCORN treats negatively correlated windows as market mirrors, and FCORN uses a sigmoid weighting function to give partially correlated windows varying influence.
Parameter experiments across stock and market-index datasets compare windows, correlation thresholds, and ensemble settings. Reported outcomes differ substantially by dataset; some configurations appear strong, while others need more parameter coverage or further analysis. The author warns that unusually high results in one historical sample may be an outlier and that selected results do not establish live-trading suitability. The approach is long-only in the described formulation, and the examples rely on historical parameter exploration.
Key ideas
- CORN selects historical market windows using Pearson correlation with the current window.
- CORN-U averages across parameterized experts, while CORN-K favors experts with stronger recent performance.
- SCORN uses negatively correlated periods as potential market mirrors.
- FCORN assigns graded importance to windows through a sigmoid activation function.
- Results and preferred parameters vary across datasets, so historical findings require further validation.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.