The paper introduces the Consensus-Bottleneck Asset Pricing Model (CB-APM), a novel interpretable deep learning framework for stock return prediction. CB-APM employs a concept-bottleneck architecture that explicitly models analyst consensus as an intermediate layer, treating professional beliefs as a sufficient statistic for the market's high-dimensional information set. The model consists of two stages: a nonlinear function that synthesizes firm characteristics and macroeconomic variables into consensus-like latent expectations, and a linear pricing stage that maps these expectations to expected returns. This design achieves interpretability-by-design, as the bottleneck constraint acts as an endogenous regularizer that improves out-of-sample predictive accuracy while anchoring inference to economically meaningful drivers. The model is estimated jointly via a weighted loss function that balances return prediction and consensus approximation, with a hyperparameter controlling the strength of the bottleneck.

Empirically, using a comprehensive dataset of U.S. stocks from 1994 to 2023, CB-APM demonstrates substantial improvements in long-horizon return prediction, achieving an out-of-sample R² of 10.46% for annual returns compared to 7.63% for a standard deep learning benchmark. Portfolio-level analyses confirm the model's economic relevance: portfolios sorted on CB-APM forecasts exhibit strongly monotonic payoff structures, with high-minus-low spreads approaching 2.3% per month for regularized specifications. Long-short portfolios achieve economically significant and stable out-of-sample performance, with mean monthly log returns rising from 1.53% to 2.20% and annualized Sharpe ratios improving from 1.10 to 1.44 as the bottleneck tightens. Furthermore, Gibbons–Ross–Shanken tests reveal that traditional factor models increasingly fail to price portfolios formed on CB-APM's predicted returns, indicating that the model uncovers priced information not spanned by linear factor models. Double sorts on model-implied returns and analysts' earnings forecasts show that the expected-return spreads are largest in states of analyst pessimism, suggesting that CB-APM effectively distills the priced component of forecasted earnings while adjusting for optimism-driven noise. These results establish CB-APM as a principled framework that integrates interpretable deep learning with asset pricing theory, offering both strong predictive performance and economic transparency.