Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Neuro-Fuzzy System for Interpretable Long-Term Stock Market Forecasting

Miha Ožbot, Igor Škrjanc, Vitomir Štruc

TL;DR

The paper addresses the challenge of achieving accurate yet interpretable long-horizon forecasting for multivariate stock data. It proposes Fuzzformer, a sequence-to-sequence neuro-fuzzy architecture that combines an LSTM encoder, Temporal Multi-Head Self-Attention, and a fuzzy inference head based on Gaussian-cluster antecedents and ARIX local models to produce $H$-step forecasts. Key contributions include the integration of deep clustering to learn antecedents, a winner-takes-all training regime for local models, and multiple regularization losses to ensure separable and balanced fuzzy rules. Empirical results on S&P 500 data with related indicators show that Fuzzformer has competitive RMSE compared to ARIMA and LSTM while offering interpretable information flow through its fuzzy components, indicating potential for practical, transparent stock-market forecasting and analysis.

Abstract

In the complex landscape of multivariate time series forecasting, achieving both accuracy and interpretability remains a significant challenge. This paper introduces the Fuzzy Transformer (Fuzzformer), a novel recurrent neural network architecture combined with multi-head self-attention and fuzzy inference systems to analyze multivariate stock market data and conduct long-term time series forecasting. The method leverages LSTM networks and temporal attention to condense multivariate data into interpretable features suitable for fuzzy inference systems. The resulting architecture offers comparable forecasting performance to conventional models such as ARIMA and LSTM while providing meaningful information flow within the network. The method was examined on the real world stock market index S\&P500. Initial results show potential for interpretable forecasting and identify current performance tradeoffs, suggesting practical application in understanding and forecasting stock market behavior.

A Neuro-Fuzzy System for Interpretable Long-Term Stock Market Forecasting

TL;DR

The paper addresses the challenge of achieving accurate yet interpretable long-horizon forecasting for multivariate stock data. It proposes Fuzzformer, a sequence-to-sequence neuro-fuzzy architecture that combines an LSTM encoder, Temporal Multi-Head Self-Attention, and a fuzzy inference head based on Gaussian-cluster antecedents and ARIX local models to produce -step forecasts. Key contributions include the integration of deep clustering to learn antecedents, a winner-takes-all training regime for local models, and multiple regularization losses to ensure separable and balanced fuzzy rules. Empirical results on S&P 500 data with related indicators show that Fuzzformer has competitive RMSE compared to ARIMA and LSTM while offering interpretable information flow through its fuzzy components, indicating potential for practical, transparent stock-market forecasting and analysis.

Abstract

In the complex landscape of multivariate time series forecasting, achieving both accuracy and interpretability remains a significant challenge. This paper introduces the Fuzzy Transformer (Fuzzformer), a novel recurrent neural network architecture combined with multi-head self-attention and fuzzy inference systems to analyze multivariate stock market data and conduct long-term time series forecasting. The method leverages LSTM networks and temporal attention to condense multivariate data into interpretable features suitable for fuzzy inference systems. The resulting architecture offers comparable forecasting performance to conventional models such as ARIMA and LSTM while providing meaningful information flow within the network. The method was examined on the real world stock market index S\&P500. Initial results show potential for interpretable forecasting and identify current performance tradeoffs, suggesting practical application in understanding and forecasting stock market behavior.

Paper Structure

This paper contains 7 sections, 11 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Temporal Multi-Head Self-Attention
  4. Fuzzy Inference System
  5. Training losses
  6. Experimentation
  7. Conclusion

Figures (2)

  • Figure 1: The Fuzzformer architecture. The orange dots at the outputs of the LSTM and Dense layers represent the dropout mechanisms. The red and blue connections illustrate the data flow of the antecedent latent space clustering samples and the local model regression variables, respectively. The variable $B$ is the batch size.
  • Figure 2: A multi-horizon forecast of the Fuzzformer with $p = 30$, test length $N {=} 60$, and horizon $H {=} 30$. The model predicts a bounce after a market drop.