How DeepAries’s Adaptive Rebalancing Timing Boosts Portfolio Returns

Background

Traditional portfolio optimization, such as Markowitz’s mean‑variance model and CAPM, relies on static, single‑period decisions and often uses heuristic periodic rebalancing, which can incur unnecessary transaction costs in stable markets or fail to react quickly to volatile conditions.

Recent reinforcement‑learning (RL) approaches (e.g., EIIE, FinRL, DeepTrader, HADAPS) treat rebalancing as a fixed‑interval task, typically daily, ignoring market dynamics and leading to suboptimal performance.

Transformer‑based time‑series models (e.g., RAT, DeepClair) have demonstrated strong ability to capture complex temporal and cross‑asset dependencies without extensive feature engineering, motivating their use in financial RL.

Problem Definition

The paper formulates adaptive portfolio management as an RL problem where the agent must simultaneously decide (1) the optimal rebalancing timing and (2) the asset allocation at each decision point.

Rebalancing timing: Choose when to rebalance to adapt to market changes.

Asset allocation: Determine portfolio weights to maximize return.

Method

DeepAries is an end‑to‑end deep RL framework that jointly learns a discrete rebalancing interval and a continuous portfolio weight vector.

Key challenges include the high‑dimensional, non‑stationary nature of market data, the limitation of fixed rebalancing intervals, and the instability of learning both discrete and continuous actions in a single policy.

Core ideas :

Explore diverse Transformer variants (classic Transformer, Informer, Reformer, Autoformer) to encode temporal and cross‑sectional dependencies.

Introduce a discrete policy that selects the next rebalancing interval from a candidate set H = {h_1, …, h_L}.

Use Proximal Policy Optimization (PPO) to jointly optimize the interval policy and a Gaussian policy for portfolio weights.

Algorithm Flow

Feature extraction: At each timestep t, historical data X(t) is fed into a Transformer encoder M_{\theta} to produce hidden representation H(t), followed by time‑attention and feed‑forward layers to obtain asset embeddings.

Adaptive interval selection: Function f_{adapt} maps the embedding e(t) to logits z^{adapt}(t); a softmax yields a probability distribution over H, from which the interval h is sampled.

Portfolio allocation: Function f_{port} generates Gaussian parameters \mu(t) and \sigma(t) for each asset; a sample gives raw weights, which are passed through tanh and normalized to obtain the final allocation w(t).

Environment transition: Using the selected interval h and weights w(t), the portfolio return R_t and value V are updated.

Reward adjustment: If the chosen interval matches the optimal interval h^*, the return is boosted by a factor b (a predefined coefficient).

Strategy optimization: The joint policy is trained with PPO, minimizing the clipped policy loss L_{PPO} plus a value loss L_{value} (mean‑squared error between predicted and target returns). The coefficient \alpha_v balances policy improvement against value‑function accuracy.

Experiments

Setup: Experiments run on four major markets (US DJ30, Europe FTSE100, Korea KOSPI, China CSI300) over 20 years. Each configuration is repeated ten times with different random seeds; the trial with the lowest validation loss is reported.

Evaluation metrics: Compound Annual Growth Rate (CAGR), Sharpe Ratio (SR), Sortino Ratio (SoR), Calmar Ratio (CR), and Maximum Drawdown (MDD).

Results:

DeepAries outperforms all baselines on 17 out of 20 metric‑market combinations, achieving the best CAGR, SR, and SoR across all markets.

In the CSI300 market, where other methods yield negative returns, DeepAries still produces positive risk‑adjusted returns.

Ablation replacing the adaptive interval component with a fixed daily rebalancing shows the adaptive approach dominates on every metric in all markets; a fixed monthly schedule matches DeepAries only in a few stable markets, suggesting market‑specific dynamics.

Combining iTransformer with the adaptive interval further improves performance; for example, in DJ30, CAGR rises from 2.31 % (fixed) to 7.90 % (adaptive), SR from 0.048 to 0.130, and MDD drops from 21.81 % to 13.67 %.

Under increasing transaction‑cost scenarios, the adaptive strategy exhibits greater robustness than the fixed‑daily baseline, mitigating cost‑induced performance degradation.