Registration & Breakfast

Opening Remarks

Relevance-Based Importance: A Comprehensive Measure of Variable Importance in Prediction

 

Mark Kritzman, CFA, CEO of Windham Capital Management

David Turkington, CFA, Senior Managing Director and Head of State Street Associates

 

The notion of variable importance is not uniquely defined. If a prediction is formed from a linear regression model, it is common to measure variable importance as a t-statistic, but a t-statistic is difficult to interpret if the predictive variables are collinear, and it is uninterpretable if the relationship between the predictive variables and the outcomes shifts as conditions change. A Shapley value measures variable importance when a prediction is formed from machine learning models. It is robust to collinearity and conditionality, but it does not account for a variable’s contribution to the reliability of individual predictions. It only considers a variable’s contribution to the reliability of predictions on average across all predictions.

 

A new measure of variable importance, called relevance-based importance, is introduced that, unlike a t-statistic, is robust to collinearity and conditionality and, unlike a Shapley value, accounts for a variable’s contribution to the reliability of individual predictions. In the special case in which the predictive variables are uncorrelated with one another and the relationship remains constant, relevance-based importance provides the same information as a t-statistic when averaged across all predictions. When relevance-based importance is averaged across all predictions, it converges to the Shapley value where the chosen value function is the R-squared of a linear regression model.

Dynamic Asset Allocation with Asset-Specific Regime Forecasts

 

John M. Mulvey, Ph.D., Professor, Bendheim Center for Finance, Center for Statistics and Machine Learning, Princeton University

 

This research introduces a novel hybrid regime identification-forecasting framework designed to enhance multi-asset portfolio construction by integrating asset-specific regime forecasts. Unlike traditional approaches that focus on broad economic regimes affecting the entire asset universe, the framework leverages both unsupervised and supervised learning to generate tailored regime forecasts for individual assets.

 

Initially, the statistical jump model, a robust unsupervised regime identification model, is used to derive regime labels for historical periods, classifying them into bullish or bearish states based on features extracted from an asset return series. Following this, a supervised gradient-boosted decision tree classifier is trained to predict these regimes using a combination of asset-specific return features and cross-asset macro-features. The framework is applied individually to each asset in the universe. Subsequently, return and risk forecasts that incorporate these regime predictions are input into Markowitz mean-variance optimization to determine optimal asset allocation weights.

 

The efficacy of this approach is demonstrated through an empirical study on a multi-asset portfolio comprising twelve risky assets, including global equity, bond, real estate, and commodity indexes spanning from 1991 to 2023. The results consistently show outperformance across various portfolio models, including minimum-variance, mean-variance, and naive-diversified portfolios, highlighting the advantages of integrating asset-specific regime forecasts into dynamic asset allocation.

Morning Break

Quantitative Models with Qualitative Scenarios: Simulation with Transformer and Large Language Models

 

Irene Aldridge, President and Head of Research at AbleMarkets

 

This research presents a novel framework for simulating financial time series by integrating the temporal fusion transformer (TFT) model. The framework incorporates scenarios derived from large language models (LLMs) to improve the adaptability and accuracy of financial simulations. By utilizing the TFT model, the framework excels in handling complex and dynamic financial data, offering enhanced explanatory power by leveraging a wide range of time-varying inputs.

 

In addition, by incorporating an LLM-based scenario generator, the framework captures both quantitative data and qualitative insights, providing a more comprehensive tool for financial analysis. To demonstrate the superior performance of the framework, stock returns of representative companies are simulated. The simulation performance is evaluated using various time-series distance metrics and financial risk management metrics. This hybrid system significantly improves upon traditional Monte Carlo simulations by producing higher-fidelity data, enabling more informed decision-making in risk management.

Networking Lunch

Thematic Investing in Total Portfolio Factor Lens

 

Wai Lee, Ph.D., Head of Systematic Research and Senior Portfolio Manager, All Spring Global Investments

 

Themes are emerging, transient, or localized factors that are often absent from conventional investment risk models. This paper applies a closed-form framework to enhance risk models by incorporating thematic characteristics that are independent of traditional factors. Using news narratives for themes identification and existing thematic exchange-traded funds as case studies, we demonstrate how the framework decomposes risks and enables portfolio construction with targeted exposures to themes that are distinct from conventional factors.

Exit Predictions for Venture Capital over Different Economic Regimes

 

Linus Franngard, Managing Director, Systematic Active Equity Portfolio Management Group, BlackRock

 

This study models the probabilities of exit for Venture Capital (VC) portfolio companies, where exits are defined by going public or being acquired. Using a feature set consisting of fundamental, sentiment, and macro-thematic predictors of VC exits, it compares differences in feature sensitivities between a full 23-year sample and a contractionary regime-based sample. The analysis finds that features like leadership size, new investors, and total equity or debt raised can positively predict exit probabilities over the long run. Relative to the full sample, statistically and economically different patterns emerge during the Dot Com Bubble Burst in 2001 and the Global Financial Crisis over 2007 to 2009. During these downturns, industry momentum contributes positively to exit probabilities, and management team experience becomes more important in predicting exits. In the Covid-19 Pandemic in 2020, exit forecasts using a model estimated during contractionary periods line up closely with the industries that most benefited from social distancing.

Afternoon Break

Optimizing Large Language Models For Sustainable Investors

 

Che Guan, Ph.D., Principal Data Scientist, AllianceBernstein L.P.

 

This study uses large language models (LLMs) and natural language processing (NLP) to extract environmental, social, and governance (ESG) insights from real-time news, creating an expert-annotated dataset to evaluate ESG classification, firm relevance, and sentiment. The fine-tuned models outperform pre-trained ones in ESG detection, firm impact, and sentiment analysis. Furthermore, in-context learning does not improve performance, indicating optimal tuning.

 

Event studies and backtests show that the sentiment signals predict underperforming stocks, with higher model confidence in negative sentiment correlating to worse outcomes. These findings emphasize the value of fine-tuning models with expert-annotated data and leveraging ESG sentiment signals to generate investment insights from qualitative data to enhance alpha generation.

Virtuous or Vain? Complexity in Equity-Premium Forecasting

 

Iro Tasitsiomi, Ph.D., Head of AI & Investments Data Science, T. Rowe Price Group Inc.

 

The recent debate on the “virtue of complexity” in equity-premium forecasting raises a practical question: when does additional flexibility generate genuine timing skill, rather than simply amplifying noise or embedding passive exposures? Using the Goyal–Welch (GW) macro-financial predictors, we reevaluate linear benchmarks, Random Fourier Feature (RFF) ridge regression, exact Gaussian RBF kernel ridge regression (KRR), and shallow neural networks within a common empirical framework. We quantify performance using both the information coefficient and the Sharpe ratio of a market-timing payoff. First, we show that the RFF expansion—at the heart of the debate—induces a covariance geometry with a large near-null space and a broad active spectrum, consistent with substantial overfitting risk. Second, in the moderate-n GW setting, exact RBF KRR is both more performant and dramatically faster as a nonlinear benchmark, calling into question the need for large RFF discretizations. Third, at similar (and sometimes lower) levels of complexity than RFFs, shallow neural networks deliver superior out-of-sample performance, although many of the top-performing networks are statistically indistinguishable from one another. Finally, using demeaned-weight payoffs and a mean–covariance decomposition, we show that strong models’ Sharpe ratios are driven primarily by covariance (timing skill), rather than by passive mean-exposure artifacts that can arise with non-centered targets and/or no-intercept specifications.

Closing Remarks and Networking