ConfFiNE: A Confusion-Based Financial Neighborhood Estimator with Directed and Validated Peer Similarity
Si-Lu Li, Si-Yu Li, John M. Mulvey
Abstract
Identifying an asset’s peers is central to relative-value trading yet difficult in cryptocurrency markets, where a dominant common factor moves all assets together and compresses pairwise correlations. We introduce ConfFiNE (Confusion-Based Financial Neighborhood Estimation), built on a natural intuition: two assets are similar when a classifier trained to recognize each asset from its recent risk–return behavior keeps mistaking one for the other. ConfFiNE offers two advantages: similarity is directional and read from held-out validation predictions, so it is validated out of sample; and the peer map, built from the classifier’s standard probability output, is independent of the trading strategy. To compare peer maps fairly, we hold a single reversal trading rule fixed, so the map is the only moving part. In walk-forward experiments on cryptocurrencies, ConfFiNE outperforms Pearson correlation with a significant Sharpe-ratio difference, attains the highest realized Sharpe ratio among sixteen peer-map estimators, and exceeds BTC buy-and-hold and the daily-rebalanced equal-weight market.
Source: semanticscholar · PDF
Read the AI summary and key takeaways for traders on WOBR Quant Research.