Inferring Fund Holdings with Extended Linear Cloning and Sequential Selection
Summary
The article describes inferring a fund’s constituents and weights from its total value time series. It compares Extended Linear Cloning (ELC), which uses regression coefficients as asset importance scores and retains the highest scoring candidates, with Sequential Oscillation Selection (SOS), which incrementally tests features and adds combinations that improve predictive fit. The experiment uses S&P 500 stocks as candidates, nine sector ETFs as target portfolios, and a 12 month price history adjusted for dividends.
SOS substantially outperforms ELC in the reported constituent classification results, including on Matthews correlation coefficient, which is useful when class labels are imbalanced. The reported evidence is limited to this experiment and its short data window. ELC tends to overestimate portfolio size; imposing minimum weights or a maximum number of holdings improves both approaches. The methods also require liquid assets with daily prices, and earlier holdings disclosures could provide useful prior information.
Key ideas
- ELC infers holdings by fitting a regression and ranking candidate assets by their coefficients.
- SOS builds candidate portfolios iteratively by testing individual assets and useful combinations.
- In the reported S&P 500 candidate universe experiment, SOS outperformed ELC on accuracy and MCC.
- MCC helps assess constituent classification when the number of included and excluded assets is unbalanced.
- Portfolio size constraints and minimum weight thresholds improved results, while daily price availability limits applicability.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.