Interpretable Company Similarity with Sparse Autoencoders

TL;DR

This work develops an interpretable framework for company similarity by applying sparse autoencoders to SEC company descriptions, producing sparse, interpretable features that are clustered with a minimum spanning tree approach. The authors show that SAE-derived clusters outperform traditional SIC/GICS codes and embeddings in capturing intra-cluster return correlations and in co-integration-based pairs trading, while enabling simple, rule-based explanations for clusters. They compare sparse-feature clustering with embedding-based methods and report strong performance gains and robust interpretability across years, including rolling-out-of-sample validation. The dataset and code release further support research into finance-focused interpretable representations of corporate descriptions and their use in risk management and trading strategies.

Abstract

Determining company similarity is a vital task in finance, underpinning risk management, hedging, and portfolio diversification. Practitioners often rely on sector and industry classifications such as SIC and GICS codes to gauge similarity, the former being used by the U.S. Securities and Exchange Commission (SEC), and the latter widely used by the investment community. Since these classifications lack granularity and need regular updating, using clusters of embeddings of company descriptions has been proposed as a potential alternative, but the lack of interpretability in token embeddings poses a significant barrier to adoption in high-stakes contexts. Sparse Autoencoders (SAEs) have shown promise in enhancing the interpretability of Large Language Models (LLMs) by decomposing Large Language Model (LLM) activations into interpretable features. Moreover, SAEs capture an LLM's internal representation of a company description, as opposed to semantic similarity alone, as is the case with embeddings. We apply SAEs to company descriptions, and obtain meaningful clusters of equities. We benchmark SAE features against SIC-codes, Industry codes, and Embeddings. Our results demonstrate that SAE features surpass sector classifications and embeddings in capturing fundamental company characteristics. This is evidenced by their superior performance in correlating logged monthly returns - a proxy for similarity - and generating higher Sharpe ratios in co-integration trading strategies, which underscores deeper fundamental similarities among companies. Finally, we verify the interpretability of our clusters, and demonstrate that sparse features form simple and interpretable explanations for our clusters.

Figures (8)

• Figure 1: Distribution of summed feature activations.
• Figure 2: Overall Mean Correlation ($\text{MC}(G_k)$) of $G_\text{CD}$ (Normalized Cosine Distance Cluster Group) vs PaLM vs SIC Benchmarks between 1996-2020. Note that we use PaLM and SIC-codes for comparison, as they have the highest $\text{MC}(G_k)$ among the embedding-based and traditional benchmark groups, respectively.
• Figure 3: Interpretability Score of Features by Percentage of Clusters ($G_\text{CD}$) where Features are Important. Data selected between 100% (all features) and 1%.
• Figure 4: Interpretability Score of Features by Percentage of Clusters ($G_\text{CDR}$) where Features are Important.
• Figure 5: Optuna Study – Histogram of Sparse Features' MST cutoff thresholds. Maximizing Threshold = -3.130.

Theorems & Definitions (1)

• proof