Predictability and Complexity Dynamics in High-Frequency Financial Machine Learning
This paper analyzes dynamic predictability and model complexity in high-frequency financial machine learning, highlighting revalidation, double ascent-descent, and estimator stability.
What it examines
Using ultra-high frequency stock data, the study examines how model complexity and re-validation affect prediction accuracy in financial machine learning. It compares linear and non-linear methods (e.g., ridge, tree-based models, neural networks) to assess dynamic predictability and the benefits of continuous hyperparameter tuning under rapid market changes.
What it concludes
Results show that non-linear models outperform simpler ones and that frequent re-estimation markedly boosts predictive accuracy. Retail trading activity and market volatility drive these dynamics. This research can enhance high-frequency trading, risk management, and adaptive investment strategies, with future studies needed to refine theoretical models.
Evidence objects
Advanced machine learning models, such as gradient boosted trees and neural networks, predict financial returns more accurately than linear models, with performance tied to market conditions and retail trading activity.
key_findings bullet 1 · key_findings · validation V0
Researchers discover a surprising 'double ascent-descent' phenomenon where adding older data initially improves predictions before deteriorating accuracy, emphasizing the urgent need for frequent re-estimation amidst evolving, significantly volatile market conditions.
key_findings bullet 2 · key_findings · validation V0
Using high-frequency trades and quotes, methods including principal component analysis, Shapley values, and random Fourier features reveal model stability, while limited sample size and market specificity mandate urgent further refinement.
key_findings bullet 3 · key_findings · validation V0
This paper presents an original exploration of high-frequency stock return predictability employing advanced machine learning techniques. Emphasizing dynamic model complexity and the novel $\text{double ascent-descent}$ phenomenon, it innovatively uses Shapley values to assess predictor importance. Its compelling insights provide fresh academic perspectives and practical relevance for quantitative finance research remarkably.
key_findings bullet 4 · key_findings · validation V0
Raw abstract and provenance
- … Second, this paper adds to an expanding literature in financial econometrics and … in financial economics that characterizes the role of retail traders in financial markets. …
Source row: 1563 · abstract type: snippet