Quantile XGBoost and SHAP in Creating and Explaining Forecasting Models for AI Tokens

This research aims to develop a model and methodology for forecasting logarithmic returns (logreturns) for a new asset class of AI (Artificial Intelligence) tokens. To obtain one day ahead forecasts for the 0.9, 0.5, and 0.1 quantiles, the use of XGBoost quantile models is proposed, which represent an ensemble, based on gradient boosted regression trees. Quantile models have an advantage in terms of forecasting over traditionally used regression models because they allow for the estimation of not only point forecasts but also of their confidence intervals, while remaining robust to outliers. This is especially important when forming the forecasts for various cryptocurrencies’ market characteristics, which are known to be highly volatile. In addition to forecasting, the study conducts a post-forecast analysis using the SHAP (Shapley Additive explanations) method, which allows to interpret the XGBoost model, revealing key factors that are important for forecasting AI tokens’ logreturns. Based on the results of feature importance analysis with SHAP, a significant influence of AI stocks’ market characteristics, cryptocurrency market investor sentiment, seasonal fluctuations, as well as features related to the Blockchain ecosystem was identified. The paper also discusses and addresses the shortcomings of modern forecasting and post-forecasting analysis approaches for time series in general. The obtained results, in addition to academic interest, are relevant for private investors, risk managers, firms and regulators.

forecasting, quantile regression, AI tokens, XGBoost, cryptocurrencies, SHAP