Beating the Clock: Volatility-Guided Deep Learning for Intraday Index Trading

Abstract

Despite the increasing popularity of algorithmic trading, most academic models primarily focus on end-of-day signals and neglect the complex challenges associated with high-frequency intraday index trading. This is particularly evident in emerging markets such as India, where over 90% of retail traders incur losses due to poor transaction timing and noncompliance with regulations. This paper introduces a volatility informed reinforcement learning system for trading on the Nifty-50 index. The model is trained on over a decade of 1-minute OHLC data. Initially, we identify statistically significant timeframes with high potential and further refine these to eliminate noise, determining the optimal times for entry. A rule-based simulator ensures realistic execution by adhering to stringent stop-loss/target logic and daily trading constraints. A Deep Q-Network (DQN) agent acquires the ability to categorize current microstructure context into Long, Short, or Hold signals within designated volatility intervals. Backtests demonstrate that the DQN agent significantly outperforms conventional indicators (such as RSI, MACD, BB, and Supertrend) and random baselines regarding return, sharpe ratio, and drawdown.