---
title: HAN-Mamba: Hierarchical Selective State Space Networks for Multi-Scale Financial Volatility Forecasting
url: https://www.ml-quant.com/papers/arxiv/2610.10323/
site: ML-Quant (https://www.ml-quant.com)
updated: 2026-10-09
license: Summaries CC BY 4.0; links go to the original sources
index: https://www.ml-quant.com/llms.txt
identifier: arXiv:2610.10323
source_url: https://arxiv.org/abs/2610.10323
featured: 2026-10-09
citations: unknown
topic: Derivatives & Volatility
---


# HAN-Mamba: Hierarchical Selective State Space Networks for Multi-Scale Financial Volatility Forecasting

Replaces transformers with selective state space encoders in a hierarchical architecture for realized volatility forecasting, reducing error to 0.1927 on the Optiver benchmark with 33 percent fewer parameters than the attention variant.

- Source: https://arxiv.org/abs/2610.10323
- Identifier: arXiv:2610.10323
- Released: 2026-10-08
- First featured: Quant Letter No. 134 (2026-10-09): https://www.ml-quant.com/issues/2026-10-09/
- Citations (Semantic Scholar): not tracked
- Published in: not yet
- Topic: Derivatives & Volatility
- Authors: Mihai Bogdan Deaconu, Ioan Daniel Pop

## Abstract (arXiv, CC0)

Short-horizon realized volatility forecasting requires the integration of market information that evolves at incompatible temporal resolutions, from second-level order book dynamics to weekly regime drift. Our conference work introduced HAN-T, a hierarchical architecture in which scale-specific Transformer encoders process short, mid, and long-horizon streams and a learned attention fuser weighs their contributions. This article replaces the quadratic attention encoders with selective state space (Mamba) encoders while retaining attention only in the fuser, where the input is a three-token set rather than a long sequence. The resulting hybrid, HAN-Mamba, summarizes each stream through a recurrent state whose input-dependent gating matches two structural properties of volatility: persistent but decaying memory and abrupt regime shifts. On the Optiver Realized Volatility Prediction benchmark under time-aware five-fold cross-validation, HAN-Mamba improves mean RMSPE over HAN-T (0.1942 vs. 0.1965) with 33% fewer parameters. Its linear-time encoders further allow the high-frequency context to be extended from 60 to 240 buckets, reducing error to 0.1927 where the attention variant saturates, and support constant-time streaming updates at inference. Ablations attribute the gains to the encoder swap, confirm that the hierarchical prior transfers across sequence-model families, and show that the permutation-invariant attention fuser remains the correct mechanism for cross-scale integration.

## Related

- [Stock Volatility Prediction Based on Transformer Model Using Mixed-Frequency Data](https://www.ml-quant.com/papers/arxiv/2309.16196/): A new model combining macroeconomic indicators, stock technical indicators, and Baidu search indices significantly improves stock volatility prediction, reducing error from 1.00 to 0.86.
- [Realised Volatility Forecasting: Machine Learning via Financial Word Embedding](https://www.ml-quant.com/papers/arxiv/2108.00480/): A new NLP framework shows that adding news text can improve stock volatility forecasting, particularly during volatile times, when combined with standard models.
- [Stochastic Volatility Modelling with LSTM Networks: A Hybrid Approach for S&P 500 Index Volatility Forecasting](https://www.ml-quant.com/papers/arxiv/2512.12250/): A combined model of Stochastic Volatility and Long Short Term Memory networks offers better volatility predictions for the S&P 500, outperforming traditional models for improved risk assessment.
- [Applying Informer for Option Pricing: A Transformer-Based Approach](https://www.ml-quant.com/papers/doi/10-5220-0013320900003890/): The research uses the Informer neural network for option pricing in financial markets, showing its improved performance over traditional models and improving financial forecasting.
- [Foundation Time-Series AI Model for Realized Volatility Forecasting](https://www.ml-quant.com/papers/arxiv/2505.11163/): The study finds that the TimesFM model, with incremental fine-tuning, is effective for volatility forecasting in financial risk management, outperforming traditional models.
- [Modeling Regime Structure and Informational Drivers of Stock Market Volatility via the Financial Chaos Index](https://www.ml-quant.com/papers/arxiv/2504.18958/): The research uses the Financial Chaos Index to study stock market volatility, identifying three market types and using sentiment predictors for volatility forecasting.