Earnings-call Q&A alignment as an equity signal

1Introduction

The information content of an earnings call extends well beyond the headline numbers and the prepared remarks. The Q&A section in particular reveals which questions management is willing or able to answer cleanly and which they evade. If a measurable signal exists in that evasion, it is the kind of edge that compounds: every quarter, every covered name, several times per year, in language that is technically already public.

2Method

2.1Embeddings

We use ProsusAI/FinBERT, a BERT model pre-trained on financial text, for question and answer embeddings. The 768-dimensional output captures domain-specific vocabulary (guidance, margins, capital allocation, tail risk) that generic BERT embeds poorly. We considered fine-tuning FinBERT itself but, with limited labelled data, the marginal gain is unlikely to justify the complexity.

2.2Contrastive architecture

The classifier learns from positive Q&A pairs (real question paired with the corresponding management answer) and negative Q&A' pairs (the same question paired with a randomly-sampled answer from a different call). Training minimises distance between aligned pairs and maximises it for misaligned ones. This is simpler than a multi-class formulation and gives a continuous alignment score that downstream backtesting can threshold flexibly.

2.3Loss

The combined loss has three terms: cross-entropy for binary aligned/misaligned classification, mean-squared error against per-pair alignment scores, and a SimCSE-inspired contrastive term to keep representations well-spread in embedding space. The trade-off is more hyperparameters; the payoff is a richer signal that is less brittle than a single-objective formulation.

2.4Backtesting

Historical alignment scores are evaluated against forward 20-day and 60-day equity returns. Reporting is by Sharpe ratio, information ratio, and drawdown analysis, with attribution by question topic (margins, guidance, competition, macro, capital structure).

3Implementation

4Results (planned)

• Alignment score correlation with forward returns: target > 0.15.
• Backtest Sharpe ratio: target > 0.8.
• Question-categorisation accuracy: target > 85%.
• Training corpus: 500+ earnings calls covering S&P 500 constituents.

5Critical path

1. Labelled training data. Hand-score 100–200 Q&A pairs with alignment labels. The classifier cannot be fit without this.
2. Train the alignment classifier on the labelled set, tuning the combined loss.
3. Backtest against 2020–2025 earnings seasons across forward 20 and 60-day windows.
4. Attribute the signal by question topic to understand what is actually driving it.
5. Expand to international equities if the US signal validates.

6Discussion

If the signal validates, this becomes a personal trading strategy. If it does not, the architecture (FinBERT plus contrastive learning plus a transparent backtest) generalises to other financial NLP problems where the question is whether language adds information beyond the numbers. Either outcome produces something useful; only the labelling work is wasted in the negative case, and even that builds expertise.

7References