📈 economics

Adaptive Position Rebalancing via Regime-Conditioned Contextual Bandits for Pre-Market Trading Optimization

This paper proposes an adaptive pre-market portfolio management framework for the Indian stock market that dynamically selects Buy, Sell, or Hold actions by combining unsupervised market regime detection with an ensemble of four contextual bandit models, demonstrating superior performance over baseline approaches in volatile conditions.

Original authors: Mutina Sathwik Apuroop, Avinash Paidi, Yamasani Purushotham Reddy, Harsha Vardhan Akili, Rolla Subrahmanyam

Published 2026-09-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mutina Sathwik Apuroop, Avinash Paidi, Yamasani Purushotham Reddy, Harsha Vardhan Akili, Rolla Subrahmanyam

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of finance, the most difficult challenge is not predicting the future, but recognizing the present. Markets are not static landscapes; they are living environments that shift between distinct moods. Sometimes they are optimistic and rising, a state traders call "bullish." Other times, fear takes hold, and prices tumble in a "bearish" slump. In between, the market often sits in a neutral state, moving sideways with no clear direction. Traditional computer models often struggle with this reality because they tend to apply a single set of rules to all situations, like a driver using the same steering technique whether the road is dry, icy, or covered in mud. When the environment changes, a rigid model often fails to adapt quickly enough, leading to costly mistakes. This is particularly true for the brief window before the stock market opens, a time when traders must decide whether to buy more shares, sell what they hold, or do nothing at all based on the latest information.

A team of researchers from India has proposed a new way to handle this volatility by teaching computers to recognize the market's current mood before making a decision. Their work focuses on the pre-market session of the Indian stock exchange, a short eight-minute window between 9:00 AM and 9:08 AM where traders adjust their portfolios before regular trading begins. Instead of relying on one giant, all-purpose algorithm, the researchers built a system that first identifies whether the market is in a bullish, bearish, or neutral regime. Once the mood is identified, the system switches to a specialized team of four different decision-making tools, each trained specifically for that type of environment. These tools work together to vote on the best action for each stock, ensuring that the strategy fits the current conditions rather than forcing a single strategy onto a changing world.

The researchers tested this approach using historical data from five major Indian companies across different industries, including banking, technology, and automotive sectors, covering a decade of market history from 2015 to 2025. They compared their new "regime-conditioned" system against standard models that do not distinguish between market moods. The results showed a clear advantage for the adaptive approach. In the simulations, the new system correctly identified the best action—whether to buy, sell, or hold—about 54 to 61 percent of the time across the different stocks. In contrast, the standard models, which used the same logic regardless of market conditions, only achieved accuracy rates between 34 and 52 percent. The adaptive system also accumulated more total profit over time and made fewer costly errors compared to an ideal strategy, suggesting that tailoring decisions to the specific market environment significantly improves performance.

The core of this success lies in the system's ability to divide and conquer. The researchers used a method called clustering to group past market days into three categories based on how prices and trading volumes behaved. When a new day arrives, the system checks the current market data against these groups to determine the regime. If the market is trending up, the system activates a specific set of four decision engines designed for bullish conditions. If the market is falling, it switches to a different set of engines trained for bearish conditions. These engines include both linear models, which look for straightforward patterns, and neural models, which can detect complex, non-linear relationships. By having a dedicated team for each mood, the system avoids the confusion that often plagues single-model approaches when the market suddenly shifts from calm to chaotic.

To make these decisions, the system looks at a wide range of technical signals, such as how much a stock price has moved recently, the average price over the last few weeks, and the volume of shares being traded. It does not try to predict the exact price of a stock tomorrow. Instead, it asks a simpler, more practical question: given the current market mood and the stock's recent behavior, is it better to increase the position, reduce it, or stay the same? The system learns from its own mistakes in a continuous loop. After each trading day, it sees how the stock actually performed and updates its internal rules to do better the next time. This allows the system to adapt to new information as it arrives, rather than relying on a static plan created months ago.

The study also highlights the limitations of this approach. The system relies entirely on historical price and volume data, meaning it cannot account for sudden, unexpected news events like a geopolitical crisis or a surprise corporate announcement that might cause a stock to gap up or down when the market opens. Additionally, the simulations assumed that trades could be executed instantly without fees or delays, which is not always the case in the real world. Despite these constraints, the findings suggest that treating market conditions as distinct regimes rather than a single continuous flow offers a more robust way to manage portfolios. By acknowledging that the rules of the game change depending on the environment, the researchers have demonstrated a path toward more stable and effective trading strategies, particularly in the critical moments before the market opens.

Technical Summary: Adaptive Position Rebalancing via Regime-Conditioned Contextual Bandits

Problem Statement

Managing stock portfolios in volatile, non-stationary financial markets presents a significant challenge for decision-making systems. Traditional machine learning models often function as static predictors, failing to adapt quickly when market conditions shift rapidly. This issue is particularly acute in the Indian stock market (NSE/BSE) during the pre-open session (9:00 AM – 9:08 AM), a short window where traders must decide whether to Buy, Sell, or Hold existing positions before regular trading begins.

The core problem addressed is the lack of an adaptive, online decision-support system that can:

Respond to continuous changes in market conditions (liquidity, volatility, sentiment).
Select discrete portfolio actions (Buy/Sell/Hold) based on the immediate market context rather than just predicting future prices.
Operate effectively within the specific constraints of the Indian pre-market call-auction mechanism.

Methodology

The authors propose a Market-Regime-Driven Contextual Bandit Framework. Instead of relying on a single global model for all market conditions, the system employs a modular architecture that adapts to the current state of the market.

1. Market Regime Identification

The framework first classifies the market into one of three regimes: Bullish, Bearish, or Neutral.

Data Source: Features are extracted from the NIFTY 50 index, including daily returns, volatility, and technical indicators.
Technique: Unsupervised K-Means clustering is used to group market observations into these three regimes. The cluster centroids are updated periodically as new data becomes available to reflect long-term shifts in market behavior.

2. Contextual Bandit Ensemble

For each identified market regime, the system maintains a dedicated ensemble of four contextual bandit models:

Linear Models: Linear Upper Confidence Bound (LinUCB) and Linear Thompson Sampling (LinTS). These are computationally efficient and suitable for linear relationships.
Neural Models: NeuralUCB and Neural Thompson Sampling (NeuralTS). These utilize neural networks to capture complex, non-linear relationships between context and reward.

3. Decision Mechanism

Context Vector ( $x_t$ ): Constructed using a rolling window of 10 trading days, incorporating technical indicators such as RSI, MACD, OBV, SMAs (5-day and 20-day), and price momentum.
Action Selection: When a decision is required, the system identifies the current regime ( $c_t$ ) and activates the corresponding specialist ensemble. The four bandit models within that ensemble independently predict an action. The final decision is determined by majority voting. In the event of a tie, a regime-specific preference rule is applied (e.g., preferring 'Buy' in a Bullish regime).
Reward Formulation: The reward is calculated using a dynamic threshold ( $\tau_t$ ) based on the average absolute price change observed up to day $t$ $t$ .
- Buy: Reward if the price increases by more than $\tau_t$ .
- Sell: Reward if the price decreases by more than $\tau_t$ .
- Hold: Reward if the price change remains within $[-\tau_t, \tau_t]$ .
- Penalty: -1 for actions that do not align with the price movement relative to the threshold.

4. Learning Process

The framework operates in a day-by-day streaming simulation:

Training Phase: Models are trained on expanding historical data blocks, updating parameters incrementally after each day's observation.
Testing Phase: A strict out-of-sample streaming evaluation is conducted where the model makes predictions for the next day using only past data, updates its policy with the realized reward, and periodically re-clusters the market states.

Key Contributions

Regime-Conditioned Architecture: The paper introduces a novel approach that combines unsupervised market regime clustering with separate contextual bandit ensembles. This allows the system to specialize in specific market environments (Bullish, Bearish, Neutral) rather than attempting to learn a single policy for all conditions.
Ensemble of Linear and Neural Bandits: By integrating both linear (LinUCB, LinTS) and neural (NeuralUCB, NeuralTS) bandit models, the framework balances computational efficiency with the ability to model non-linear market dynamics.
Pre-Market Optimization: The study specifically targets the under-researched domain of the Indian pre-market session (9:00–9:08 AM), formulating it as a sequential decision-making task for position rebalancing.
Dynamic Reward Mechanism: The use of a volatility-adjusted dynamic threshold for reward calculation allows the system to adapt its sensitivity to market movements, rather than relying on fixed percentage changes.

Experimental Results

The framework was evaluated on selected Indian sectoral stocks (Hindustan Unilever, HDFC Bank, Infosys, Larsen & Toubro, Mahindra & Mahindra) using data from January 2015 to July 2025. The out-of-sample evaluation covered July 2024 to July 2025.

Accuracy: The proposed regime-conditioned framework achieved significantly higher trading decision accuracy (ranging from 54.8% to 61.2%) compared to standalone baseline models (which ranged from 34.8% to 52.0%).
Cumulative Reward: The proposed approach demonstrated a more consistent increase in cumulative reward over time, outperforming all baseline models.
Regret Analysis: The framework exhibited a lower rate of cumulative tracking regret, indicating that it maintained more consistent decisions closer to an "oracle" policy (the best possible action) even as market conditions changed.
Stability: The regime-based approach reduced the "policy drift" observed in global baseline models when markets transitioned between different states.

Significance and Claims

The paper claims that financial non-stationarity significantly degrades the performance of single global bandit models. By conditioning the decision process on identified market regimes, the proposed framework achieves:

Improved Stability: Separate models for different regimes allow the system to learn simpler reward structures specific to that environment.
Adaptive Efficiency: The system can adapt more effectively to environmental changes by switching between specialized policies.
Practical Utility: It provides a viable, data-driven decision-support tool for pre-market portfolio management in the Indian context, addressing the specific constraints of the short pre-open session.

The authors acknowledge limitations, including the reliance on historical endogenous data (lack of external event integration) and the assumption of frictionless execution (ignoring transaction costs and slippage). However, they assert that the framework offers a robust benchmark for adaptive pre-market trading strategies.