Artificial Intelligence 12 min read

AB‑SSM: Adaptive Bidirectional State‑Space Model for High‑Frequency Portfolio Management

The paper introduces AB‑SSM, an adaptive bidirectional state‑space model that incorporates a time‑varying linear structure and a bidirectional layer to capture market non‑stationarity and asset correlations, and demonstrates through extensive US, China, and crypto experiments that it outperforms traditional, deep‑learning, and DRL baselines in profit‑risk trade‑offs, efficiency, and scalability.

Bighead's Algorithm Notes

Mar 5, 2026

AB‑SSM: Adaptive Bidirectional State‑Space Model for High‑Frequency Portfolio Management

Background

State‑space models (SSMs) achieve strong long‑sequence modeling but in high‑frequency portfolio management they encounter two issues: market non‑stationarity makes linear time‑invariant SSMs unable to capture domain shifts, and the unidirectional nature prevents modeling asset correlations.

Problem Definition

High‑frequency portfolio management requires (1) handling non‑stationarity to exploit short‑lived arbitrage, (2) capturing inter‑asset correlations for allocation decisions, and (3) providing low‑latency sequential allocation.

Method

AB‑SSM comprises three modules: a sequence‑information network, a correlation‑information network, and a decision module.

Adaptive State‑Space Layer (ASSL)

ASSL replaces the fixed state‑transition matrix with an input‑dependent matrix, converting the SSM into a linear time‑varying system. The continuous‑time formulation is

After discretization the recursion becomes

This design increases sensitivity to subtle financial dynamics.

Initialization The transition matrix A is initialized with a HiPPO‑N matrix to obtain richer sequential representations:

Computation Because the time‑varying structure breaks convolutional equivalence, a parallel‑scan algorithm is used for efficient computation, and a multi‑input‑multi‑output (MIMO) SSM processes large‑scale multivariate series, reducing time‑space overhead.

Bidirectional State‑Space Layer (BSSL)

Two independent ASSLs process the input forward and backward; their outputs are concatenated to form a bidirectional representation:

A final linear layer fuses the concatenated features.

Decision Module

The module concatenates the previous portfolio vector with the market embedding from the correlation network to obtain an enhanced embedding e_{1:m}. A fully‑connected layer followed by a Softmax yields the portfolio vector; a fixed zero‑cash bias is added before Softmax:

Experiments

Setup

Datasets: real‑world data from US DJIA constituents, China SSE‑50 constituents, and five major cryptocurrencies (BTC, ETH, etc.).

Baselines: traditional methods (Market, EG, ONS, Anticor), deep‑learning methods (ALSTM, ADARNN), and deep‑reinforcement‑learning methods (EIIE, SARL, DeepTrader, CITrader).

Metrics: Annualized Return (ARR), Annualized Sharpe Ratio (ASR), Maximum Drawdown (MDD), and Annualized Volatility (AVol).

Results

Profitability AB‑SSM achieves the highest ARR and ASR and the lowest MDD on all three datasets, indicating superior profit generation and risk control. Its AVol is not the best but remains acceptable.

Portfolio value The portfolio‑value curve shows stable, strong growth across expansion, stagnation, and recession market regimes, accumulating returns above 150%.

Efficiency On the Crypto dataset AB‑SSM requires less computation time and memory than DeepTrader and CITrader, and its scalability improves with longer sequences.

Ablation study Removing any of the three key components (adaptive linear time‑varying structure, ASSL, BSSL) degrades performance, confirming each module’s contribution.