Out-of-Sample Exchange Rate Prediction: A Machine Learning Perspective (with Ilias Filippou, Mark P. Taylor, and Guofu Zhou)

Abstract: We establish the out-of-sample predictability of monthly exchange rates via machine learning techniques based on 70 predictors capturing country characteristics, global variables, and their interactions. To better guard against overfitting in our high-dimensional and noisy data environment, we adjust “off-the-shelf” implementations of machine learning techniques to induce adequate shrinkage. The resulting forecasts consistently outperform the no-change benchmark, which has proven difficult to beat. Variable importance analysis indicates that country characteristics are important for forecasting, once they interact with global variables. Machine learning forecasts also markedly improve the performance of a carry trade portfolio, especially since the Global Financial Crisis.

[Internet Appendix | SSRN link | Presented at the 2021 Vienna Symposium on Foreign Exchange Markets | Presented at the 2020 Wolfe Virtual Global Quantitative & Macro Investment Conference]

Expected Stock Returns and Firm Characteristics: E-LASSO, Assessment, and Implications (with Yufeng Han, Ai He, and Guofu Zhou)

Abstract: We develop new methods for constructing and analyzing cross-sectional stock return forecasts. We propose an E-LASSO approach that uses the LASSO, forecast combination, and forecast encompassing to implement shrinkage in a flexible manner designed to handle a large number of firm characteristics. We provide a cross-sectional out-of-sample R-squared statistic for assessing the accuracy of cross-sectional forecasts. Empirically, with presently the largest set of 193 firm characteristics, we find that our E-LASSO forecast produces significant cross-sectional out-of-sample R-squared gains and generates substantial economic value consistently over time. We further find that many firm characteristics matter, instead of a dozen or so.

[SSRN link | Notes on Frisch-Waugh-Lovell theorem in deviation form for OLS/WLS estimation of cross-sectional univariate regressions | Presented at the AFA 2019 Annual Meeting]

Sparse Macro Factors (with Guofu Zhou)

Abstract: We use machine-learning techniques to estimate sparse principal components (PCs) for 120 monthly macroeconomic variables from the FRED-MD database. Each sparse PC is a sparse linear combination of the underlying macroeconomic variables, allowing for their economic interpretation. Innovations to the sparse PCs constitute a set of sparse macro factors. Robust tests indicate that sparse macro factors corresponding to yields and housing earn statistically and economically significant risk premia. A three-factor model comprised of the market factor and mimicking portfolio returns for the yields and housing factors performs well compared to leading multifactor models in explaining numerous anomalies.

[SSRN link | Data for new sparse macro three-factor model | Best Paper Prize, INQUIRE UK & Europe Spring 2019 Residential Joint Conference]

Boosting Cryptocurrency Return Prediction (with Ilias Filippou and Christoffer Thimsen)

Abstract: We use boosted decision trees to generate daily out-of-sample forecasts of excess returns for Bitcoin and Ethereum, the two best-known and largest cryptocurrencies. The decision trees incorporate information from 39 predictors, including variables relating to cryptocurrency fundamentals, technical indicators, Google Trends searches, Reddit comments, and articles from Factiva. We use the XGBoost algorithm to boost trees and find that excess return forecasts based on boosted trees produce statistically and economically significant out-of-sample gains. We explore the importance of individual predictors and nonlinearities in the fitted boosted trees. We find that a broad array of predictors are relevant for forecasting daily cryptocurrency returns and that strong nonlinearities characterize the predictive relationships.

[SSRN link]

Bayesian Estimation of Macro-Finance DSGE Models with Stochastic Volatility (with Fei Tan)
Revise & resubmit to the Journal of Applied Econometrics)

Abstract: We develop a Bayesian Markov chain Monte Carlo algorithm for estimating risk premia in dynamic stochastic general equilibrium (DSGE) models with stochastic volatility. Our approach is fully Bayesian and employs an affine solution strategy that makes estimation of large-scale DSGE models computationally feasible. We use our algorithm to estimate the US equity risk premium in a DSGE model that includes time-preference, technology, investment, and volatility shocks. Time-preference and technology shocks are primarily responsible for the sizable equity risk premium in the estimated DSGE model. The estimated historical stochastic volatility and equity risk premium series display pronounced countercyclical fluctuations.

[Internet Appendix | MATLAB toolbox | SSRN link]

OTHER ITEMS

Forecasting Asset Returns: The State of the Art
Invited Lectures at CEMA/CUFE on July 12-13, 2016

[Slides - Lecture 1 | Slides - Lecture 2]

Forecasting Asset Returns in Realistic Environments
Invited Presentation at the CFA Montreal Asset Management Forum on October 8, 2015 (other presenters: Andrew Ang, Mark Carhart, Craig Bodenstab, Philip Tetlock)

[Slides | CFA Research Foundation Brief: Portfolio Structuring and the Value of Forecasting]