Multi-Horizon Echo State Network Prediction of Intraday Stock Returns

Giovanni Ballarin, James Robert Capra, Πέτρος ΔελλαπόρταςarXiv 2025 · DOI · arXiv

TL;DR

This paper introduces a high-efficiency framework for predicting intraday stock returns across multiple time horizons using Echo State Networks (ESNs). By leveraging "reservoir computing"—where the internal neural network weights are random and fixed—the authors achieve significant forecasting gains over linear models with a fraction of the computational cost required by traditional deep learning.

What Problem Does This Paper Address?

Predicting stock returns is the "Holy Grail" of quantitative finance, but it is notoriously difficult due to the low signal-to-noise ratio of financial data. While modern deep learning (like LSTMs or Transformers) can capture complex non-linear patterns, these models are often "black boxes" that are incredibly expensive to train and slow to run in high-frequency environments.

Conversely, simple linear models are fast but fail to capture the intricate, time-dependent relationships inherent in market data. The authors seek a "middle ground": a model that possesses the expressivity of recurrent neural networks but maintains the computational efficiency of linear regression.

Methods

The authors utilize Echo State Networks (ESNs), a subset of Reservoir Computing. The core architecture consists of:

  1. The Reservoir: A large recurrent layer where the internal connection weights ($A$) and input weights ($C$) are randomly sampled and never trained. This reservoir acts as a non-linear dynamical system that maps input signals into a high-dimensional "state space."
  2. The Readout: A simple linear layer that maps the reservoir's state to the predicted return. Only this layer is trained, typically using Ridge Regression (squared loss with $L_2$ regularization).
  3. Signal Construction: Instead of raw prices, the model ingests "z-score" signals derived from a combination of Principal Component Analysis (PCA) to remove market-wide factors and an Ornstein-Uhlenbeck process to model idiosyncratic stock residuals.

The model is "multi-horizon," meaning it is tuned to predict returns at 10-minute, 30-minute, 1-hour, 2-hour, and End-of-Day (EOD) intervals.

Datasets

The study uses high-quality intraday data from AlgoSeek:

  • Universe: 500 US stocks.
  • Resolution: 1-minute OHLC (Open, High, Low, Close) bars, downsampled to 10-minute intervals for the main experiments.
  • Timeframe: The full year of 2013 is used for evaluation, with the final quarter of 2012 used for hyperparameter tuning (via the Optuna library).

Key Insights and Ideas

  • The Virtue of Randomness: The paper leans into the "Lottery Ticket Hypothesis"—the idea that within a large enough random network, there exists a sub-network that is exceptionally well-suited for a task even before training. By keeping the reservoir random, the ESN avoids the "vanishing gradient" problems of standard RNNs.
  • Non-linear Memory: Unlike linear models that only look at a fixed window of past data, the ESN’s recurrent state equation allows it to maintain a "fading memory" of past inputs, effectively creating a non-linear summary of the stock's history.
  • Efficiency as a Feature: The authors demonstrate that their entire forecasting pipeline for 500 stocks over a full year can be run in "a handful of minutes." This makes it viable for real-time systematic trading where models must be re-estimated frequently.

Key Results

  • Superior Accuracy: The ESN model consistently outperformed both a simple linear baseline and a regularized linear benchmark.
  • Horizon Performance: The strongest gains were seen at shorter horizons. At the 10-minute horizon, the ESN achieved a 0.87% reduction in Mean Squared Forecasting Error (MSFE).
  • Statistical Significance: Using Diebold-Mariano and Model Confidence Set (MCS) tests, the authors confirmed that the ESN's improvements were statistically significant at almost all horizons, except for EOD returns where the advantage was marginal.
  • Robustness: The results were remarkably stable across different random seeds. Even though the internal weights are random, the aggregate performance of the "reservoir" is highly consistent.

Limitations and Future Work

  • Linear Readout: The authors constrained the output layer to be linear for efficiency. They suggest that adding a "shallow" non-linear output layer might further boost performance.
  • Model Ensembles: The current work uses a single ESN. The authors note that "ensembling" (averaging predictions from multiple ESNs with different random initializations) is a promising path for future research to further reduce variance.
  • Static Hyperparameters: While the readout is trained daily, the core ESN hyperparameters (like the "leak rate" or "spectral radius") were fixed based on 2012 data. Dynamic, online tuning of these parameters could potentially adapt better to changing market regimes.

Summary updated

Abstract

Stock return prediction is a problem that has received much attention in the finance literature. In recent years, sophisticated machine learning methods have been shown to perform significantly better than ''classical'' prediction techniques. One downside of these approaches is that they are often very expensive to implement, for both training and inference, because of their high complexity. We propose a return prediction framework for intraday returns at multiple horizons based on Echo State Network (ESN) models, wherein a large portion of parameters are drawn at random and never trained. We show that this approach enjoys the benefits of recurrent neural network expressivity, inherently efficient implementation, and strong forecasting performance.
BibTeX
@misc{ballarin2025multi,
  author = {Giovanni Ballarin and James Robert Capra and Πέτρος Δελλαπόρτας},
  title = {{Multi-Horizon Echo State Network Prediction of Intraday Stock Returns}},
  year = {2025},
  eprint = {2504.19623},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2504.19623},
  howpublished = {arXiv},
}
Figure