‘Bay Street Meets Machine Learning: Predicting Stock Risk Premium’

December 12, 2025

The winners of the 2025 Hillsdale Investment Management – CFA Society Toronto Research Award are Najah Attig, professor of finance and chair of the department of finance, Dalhousie University, and Chahine Attig, data science engineering student at École Nationale de la Statistique et de l’Analyse de l’Information (ENSAI), France, for their research paper titled Bay Street Meets Machine Learning: Predicting Stock Risk Premium.

Objectives of award winners’ research

The winning paper provides one of the first comprehensive academic evaluations of machine learning techniques for forecasting equity risk premiums in Canadian capital markets. Compared to U.S., European and Chinese markets, Canada has seen limited application, despite the distinct structural, liquidity and informational characteristics in the Canadian market. Markets in Canada are dominated by smallcap and value stocks with greater information asymmetry, market frictions and liquidity constraints, which pose unique challenges and opportunities for predicting returns. These highlight the need for advanced modelling approaches beyond traditional linear methods.

“This winning paper shows some promise in applying machine learning methods to extract higher dimensional signals from the Canadian stock market,” – Chris Guthrie

The researchers address two key questions:

Can machine learning models improve Canadian stock return forecasts compared to classical linear benchmarks?
Do patterns in anonymous trading – where traders’ identities are concealed to prevent information leakage or market speculation – and variations in brokers’ anonymous trading activity help predict stock returns?

Findings of study

The researchers’ key findings are:

Moderately flexible, non-linear models, especially XGBoost¹ and mid-depth neural networks, outperform deeper neural networks and traditional linear models in forecasting Canadian stock returns.
Monthly long-short portfolios based on machine learning forecasts generate average excess returns exceeding one per cent, with Sharpe ratios above 0.75, more than double the linear benchmark’s Sharpe ratio (about 0.30). Annualized returns exceed 20 per cent, with Sharpe ratio approaching 0.9, demonstrating substantial economic gains.
Predictability is strongest in small-cap and value stocks, which dominate the Canadian equity market and exhibit frictions such as low liquidity and informational inefficiencies. Machine learning models effectively capture persistent alpha opportunities in these segments.
Stocks with elevated net anonymous buying and greater broker-level order flow dispersion show stronger return predictability. Including these variables improves monthly portfolio returns by 60 to 80 basis points with Sharpe ratios near or above 0.9 when combined with flexible machine learning learners. This highlights the value of integrating informed trading proxies into forecasting.
Liquidity, bid-ask spreads (an informational asymmetry proxy) and trading volume influence forecast accuracy. Small-cap stocks, which are more affected than large-cap stocks by microstructure inefficiencies and fast-moving fundamentals, benefit most from non-linear machine learning models. Large-cap returns are driven more by systemic risk and capital structure, where regularized linear models suffice.
Penalized linear models excel in liquid, large-cap universes by providing stable, interpretable forecasts, while illiquid, less efficient segments gain from non-linear methods like gradient boosting and neural networks. This nuanced insight guides model selection in practice.

Implications for practitioners

For practitioners, the study confirms that machine learning techniques provide powerful tools to overcome limitations of

traditional forecasting models, particularly under constraints such as long-only mandates common in Canadian funds. Machine learning’s ability to uncover non-linearities and complex predictor interactions is critical in a market characterized by data complexity and frictions.

The researchers’ evaluation emphasizes out-of-sample performance and economic relevance through portfolio back tests, aligning with industry standards for robust, implementable forecasting solutions. Their findings advocate for a segmented approach to machine learning model development: simple, regularized linear models are effective in liquid, large-cap stocks, while sophisticated machine learning models are better suited to small-cap, value and less efficient segments.

Incorporating microstructure data, especially anonymous order flow metrics, emphasizes the growing importance of alternative data sources. As financial markets become increasingly complex and data-rich, non-traditional inputs will become central to adaptive, evidence-based investment decision-making.

Conclusions

The authors highlight that machine learning is emerging as a powerful technology that can enhance Canadian equity investing by extracting economically meaningful signals overlooked by traditional models.

Machine learning presents a promising frontier for improving return predictability and portfolio performance in Canadian capital markets. By exploiting rich firm-level data, macroeconomic variables and novel microstructure signals, machine learning models appear to deliver substantial gains over traditional forecasting methods.

Predictability is most pronounced in small-cap and value stocks, where market frictions are significant, and anonymous trading imbalances reveal unique insights into informed trading behaviour. The authors encourage academics and practitioners to embrace machine learning not only for forecasting improvements but also for deepening understanding of market mechanisms and investor behaviour. Advancing machine learning interpretability tools to better connect data-driven insights with economic theory will be crucial for developing robust and adaptive investment strategies in

Canada’s evolving market environment and beyond.

The researchers note that machine learning is not a substitute for economic theory but a complementary tool, particularly useful in high-dimensional, high-noise settings where conventional models struggle. While their analysis focuses on the Canadian equity market, the insights may also apply to other contexts, including emerging markets with greater structural inefficiencies and microstructure frictions. As financial data grows more complex and heterogeneous, machine learning offers a useful addition to both academic asset pricing research and applied investment strategy.

More broadly, the researchers’ findings contribute to the discussion on the integration of machine learning in addressing emerging questions in finance, including climate risk pricing, ESG-related mispricing and the role of retail investor sentiment and algorithmic trading in return dynamics. They highlight that future research should focus not only on improving predictive accuracy but also on interpreting machine learning-generated signals and linking them to underlying economic mechanisms. Such efforts can help bridge the gap between data driven inference and theory-consistent modelling, supporting a more rigorous and informed application of machine learning techniques in financial economics.

As financial markets become increasingly complex and data-rich, non-traditional inputs will become central to adaptive, evidence-based investment decision-making.

The award judging panel commented, “Running a horse race of various machine learning techniques, the winning paper finds that moderately flexible machine learning models consistently outperform deeper architectures and linear models in enhancing return forecasting and portfolio construction in Canadian equity markets. The judging team commends the authors on the fact that the research is actionable, makes use of the latest numerical techniques and that the analyses were performed in a Canadian context.”

Chris Guthrie, CEO of Hillsdale Investment Management, commented, “Applications of machine learning methods to portfolio management and investments have grown in popularity in both academia and industry, but very few have focused on Canadian markets. This winning paper shows some promise in applying machine learning methods to extract higher-dimensional signals from the Canadian stock market.”

The Hillsdale Investment Management – CFA Society Toronto Investment Research Award is open to researchers globally who conduct research related to Canadian capital markets, including both academics (e.g., professors and students) and practitioners. Author(s) of the winning research paper are awarded C$10,000. Research papers are reviewed by a panel of CFA charterholding investment experts to ensure they align with the rigorous values and standards embodied in the CFA designation.

¹ XGBoost is an open-source machine learning library based on the gradient boosting algorithm that excels at classification, regression and ranking tasks. (Definition provided by AI Overviews.)

Date December 12, 2025

Tags Hillsdale