Paper: arXiv 2308.08550

Authors: Damien Challet, Vincent Ragel

Abstract

We extend recurrent neural networks to include several flexible timescales for each dimension of their output, which mechanically improves their abilities to account for processes with long memory or with highly disparate time scales. We compare the ability of vanilla and extended long short term memory networks (LSTMs) to predict asset price volatility, known to have a long memory. Generally, the number of epochs needed to train extended LSTMs is divided by two, while the variation of validation and test losses among models with the same hyperparameters is much smaller. We also show that the model with the smallest validation loss systemically outperforms rough volatility predictions by about 20% when trained and tested on a dataset with multiple time series.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents advanced mathematical modifications to RNNs (multiple timescale approximations, GRU/LSTM architecture derivations) while rigorously evaluating performance through a systematic backtest on financial data, including hyperparameter sweeps, train/validation/test splits, and comparison against a standard benchmark (rough volatility).

Research Flowchart

  flowchart TD
  A["Research Goal: Improve RNNs for Volatility Prediction<br>with Long Memory"] --> B["Methodology: Extended LSTM with<br>Flexible Timescales"]
  B --> C["Data: Multi-Time Series Equities Dataset"]
  C --> D["Computational Process:<br>Train Vanilla vs. Extended LSTMs"]
  D --> E{"Evaluation & Comparison"}
  E --> F["Outcome 1: 2x Faster Convergence<br>& Reduced Model Variance"]
  E --> G["Outcome 2: ~20% Better Accuracy<br>than Rough Volatility Models"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes extending recurrent neural networks with several flexible timescales per output dimension to improve long‑memory prediction.

Source sentence
We extend recurrent neural networks to include several flexible timescales for each dimension of their output, which mechanically improves their abilities to account for processes with long memory or with highly disparate time scales.

Assumes 2

  • Asset price volatility exhibits long memory. model
    Source sentence
    We compare the ability of vanilla and extended long short term memory networks (LSTMs) to predict asset price volatility, known to have a long memory.
  • Training and testing are performed on a dataset containing multiple time series. data
    Source sentence
    We also show that the model with the smallest validation loss systemically outperforms rough volatility predictions by about 20% when trained and tested on a dataset with multiple time series.

Data

No dataset named in the text.

Costs

Not applicable The paper reports no trading or execution result.

Finds 4

  • Extended LSTMs require roughly half the training epochs compared to standard LSTMs. performance
    Source sentence
    Generally, the number of epochs needed to train extended LSTMs is divided by two, while the variation of validation and test losses among models with the same hyperparameters is much smaller.
  • Models with flexible timescales show much smaller variation in validation and test losses. statistical
    Source sentence
    while the variation of validation and test losses among models with the same hyperparameters is much smaller.
  • The best‑validation‑loss model outperforms rough volatility predictions by about 20%. predictive-accuracy
    Source sentence
    We also show that the model with the smallest validation loss systemically outperforms rough volatility predictions by about 20% when trained and tested on a dataset with multiple time series.
  • Adding a flexible kernel structure to LSTMs improves convergence time, overall prediction accuracy, and reduces accuracy variation across models. performance
    Source sentence
    Adding an explicit but flexible kernel structure to LSTMs brings significant improvements in every metric: number of epochs needed to reach convergence, overall prediction accuracy, and accuracy variation between models at fixed.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 33 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

https://github.com/damienchallet/LaSTM — the authors' own repository

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.