[Paper Review] FinAnchor: Aligned Multi-Model Representations for Financial Prediction

Zirui He, Hanyu Zhang+3 more|arXiv (Cornell University)|Feb 24, 2026

Stock Market Forecasting Methods0 citations

TL;DR

FinAnchor aligns embeddings from multiple frozen LLMs into a common anchor space via linear mappings, aggregates them, and trains a lightweight readout, improving financial prediction across long documents without fine-tuning base models.

ABSTRACT

Financial prediction from long documents involves significant challenges, as actionable signals are often sparse and obscured by noise, and the optimal LLM for generating embeddings varies across tasks and time periods. In this paper, we propose FinAnchor(Financial Anchored Representations), a lightweight framework that integrates embeddings from multiple LLMs without fine-tuning the underlying models. FinAnchor addresses the incompatibility of feature spaces by selecting an anchor embedding space and learning linear mappings to align representations from other models into this anchor. These aligned features are then aggregated to form a unified representation for downstream prediction. Across multiple financial NLP tasks, FinAnchor consistently outperforms strong single-model baselines and standard ensemble methods, demonstrating the effectiveness of anchoring heterogeneous representations for robust financial prediction.

Motivation & Objective

• Motivate robust financial prediction from long-form documents despite sparse actionable signals and model bias.
• Propose a lightweight multi-encoder alignment framework that avoids fine-tuning base models.
• Show that aligning heterogeneous representations yields a unified, effective downstream predictor.
• Provide interpretability analyses to understand when and why alignment improves performance.

Proposed method

• Select an anchor LLM embedding space and learn linear ridge mappings from other encoders into this anchor space.
• Standardize features and fit ridge regression mappings using training data only (no intercept).
• Aggregate aligned representations in the anchor space (e.g., compute mean across encoders).
• Train a lightweight downstream readout (MLP) on top of the aggregated anchor representation.
• Evaluate on multiple financial NLP tasks with time-based splits to assess out-of-time performance.

Figure 1: (a) Three different LLM independently predict next-day stock movement from the same earnings call transcript: Models A and B predict negative return, while Model C predicts positive return. Highlighted phrases denote text snippets deemed salient by each model. In this case, Model A and B e

Experimental results

Research questions

• RQ1 Can FinAnchor improve predictive performance over single-encoder baselines and naive ensembles across diverse financial text tasks?
• RQ2 Do aligned multi-model representations provide systematic, not random, gains evidenced by error patterns and confidence shifts?
• RQ3 Is the framework interpretable, revealing how and which evidence from different models drives predictions?

Key findings

• FinAnchor achieves the best overall performance across five datasets, outperforming strongest single-model baselines.
• Alignment reduces false positives more than false negatives, indicating improved risk control in predictions.
• Error overlaps between base models are non-trivial, enabling corrections via alignment.
• Alignment yields directionally positive confidence shifts on correctly corrected cases.
• Linear alignment quality (R^2) is high for source-to-anchor mappings, suggesting compatible geometric structures across encoder spaces.
• Case studies show evidence reweighting amplifies financially salient cues that align with ground truth.

Figure 2: Decision transitions after alignment on Stock Movement Prediction task. Counts of label transitions from Gemma to the FinAnchor under validation-chosen thresholds. The FinAnchor corrects a substantial number of false positives (FP $\rightarrow$ TN), while also introducing smaller regressio