Temporal Mixture Density Networks for Enhanced Investment Modeling
37 Pages Posted: 15 Apr 2024 Last revised: 3 Sep 2024
Date Written: April 3, 2024
Abstract
This research harnesses advancements in deep learning(DL) and statistical analysis to present a novel approach that focuses on employing a mixture model for the prediction of asset price returns and their uncertainties. This methodology distinguishes between days of extreme and normal returns, with significant enhancements provided by Mixture Density Network (MDN) methodologies.
Moreover, the study evaluates the efficacy of deep neural networks based on Long Short-Term Memory (LSTM) in modeling financial time series, we propose LSTM-Gaussian-t Network (GTN) models and compare it with Gaussian Mixture Network (GMN) which was first suggested by Nikolaev et al. (2013) based on our proposed formulation and implementation. The result shows that the Preliminary Model selection and Investment Model evaluation are utilized for the appraisal of models and investment strategies. The evaluation process initiates with three distinct criteria: the minimal validation loss (corresponding to the most negative log-likelihood), the maximal Sharpe ratio, and the optimal F1 score. Following this, the investment strategies formulated from the parameters of the selected models are implemented. The selection process culminates with the investment strategy and model that achieves the highest Sharpe ratio during the validation phase. The findings reveal that the investment strategies, informed by the predictive model, significantly augment both risk-adjusted and total returns for investors engaging in single-stock investments over a comprehensively tested investment period.
Keywords: Mixture Density Network; LSTM; Asset Price Prediction; Investment Strategies; Sharpe Ratio; F1 Score; Drawdown., LSTM
Suggested Citation: Suggested Citation