What this approach can—and cannot—do
Combining reinforcement learning (RL) with technical indicators can help structure a trading research process, but it does not guarantee profits. An RL agent learns a policy for choosing actions under uncertainty; it does not discover a reliable market edge merely because indicators are included as inputs. Indian equities are affected by liquidity, circuit limits, corporate actions, sector news, commodity prices, interest rates, and execution constraints that historical simulations can easily miss.
Use the method as an experimental decision system, not as an autonomous promise of returns. Begin with liquid, clearly defined stocks and validate every result after brokerage, exchange fees, securities transaction tax, GST, stamp duty, slippage, and taxes are considered.
Define the Gujarat chemical universe
“Gujarat chemical stock market” is not a single exchange or index. It is better treated as a research universe of chemical and specialty-chemical companies listed on NSE or BSE and operating from, or strongly associated with, Gujarat. Define the universe before collecting data.
Document:
- The exact symbols, exchange, sector classification, and listing history.
- Whether the strategy is long-only, long/short, or cash-inclusive.
- Minimum price, average traded value, and turnover filters.
- Rebalancing frequency and maximum number of simultaneous positions.
- Treatment of suspended, delisted, newly listed, or renamed companies.
Avoid survivorship bias by retaining historical constituents and delisted securities where possible. A strategy tested only on today’s successful companies will usually look stronger than it could have been in live trading.
Build clean, time-aligned data
At minimum, collect adjusted OHLCV data, corporate actions, and a trading calendar. For a serious study, add delivery data, bid–ask estimates, benchmark returns, sector data, and relevant commodity or currency variables. Chemical companies can be sensitive to crude-linked inputs, natural gas, freight, export demand, and the Indian rupee, so a price-only model may omit important context.
Calculate indicators using information available at the decision time. If the agent trades at the next session’s open, indicators must be based on the prior session’s completed data. Never use a day’s closing price to simulate a trade that supposedly occurred earlier that same day.
Useful features include:
- Trend: 10-, 20-, 50-, and 200-session moving-average distance.
- Momentum: RSI, rate of change, and MACD histogram.
- Volatility: ATR, rolling standard deviation, and Bollinger-band width.
- Liquidity: rolling traded value, volume change, and estimated spread.
- Market context: Nifty or sector benchmark return, regime volatility, and market breadth.
Do not add every indicator available. Highly correlated features increase complexity without necessarily adding information. A compact feature set is easier to audit and less prone to overfitting. Teams new to this workflow can practise the surrounding engineering fundamentals through machine learning portfolio projects for beginners in India.
Design the RL environment explicitly
Represent each decision as a state, action, transition, and reward:
- State: normalized technical features, current position, cash, portfolio value, recent returns, and risk exposure.
- Action: discrete choices such as buy, hold, and sell, or a continuous target position between zero and a defined maximum.
- Transition: the portfolio changes after the selected action and the next market observation arrives.
- Reward: risk-adjusted change in portfolio value after trading costs.
A simple reward based only on raw profit encourages unstable behaviour and excessive turnover. A more useful formulation can subtract estimated costs and penalties:
reward = portfolio_return - transaction_cost - slippage - risk_penalty
Risk penalties may reflect drawdown, volatility, concentration, or exposure beyond a predefined limit. Keep the reward interpretable. If the reward contains too many manually tuned terms, it becomes difficult to know what the agent has actually learned.
Use position sizing and hard safety rules outside the policy where appropriate. For example, cap single-stock exposure, block trades when liquidity falls below a threshold, and impose a portfolio-level drawdown stop. These controls should not be removed simply because the backtest looks profitable.
Select an algorithm and establish baselines
Start with a non-RL benchmark before using deep learning. Compare against buy-and-hold, equal-weight rebalancing, a moving-average rule, and a supervised model that predicts returns or direction. If RL cannot beat simple baselines after costs and risk adjustment, adding a more complex algorithm is unlikely to solve the problem.
Algorithm choices depend on the action space:
- DQN: suitable for small, discrete action spaces, but sensitive to non-stationary data.
- PPO: a practical starting point for policy learning and bounded position decisions.
- SAC or TD3: relevant when actions represent continuous portfolio weights, though they require careful tuning.
Use fixed random seeds, versioned configurations, and multiple training runs. Reproducibility matters more than a single impressive equity curve. The same data pipeline and evaluation discipline used in scalable machine learning infrastructure for developers can help keep experiments traceable.
Train with walk-forward evaluation
Do not randomly shuffle time-series observations. Split data chronologically into training, validation, and test periods. A walk-forward design is more realistic:
1. Train on an initial historical window.
2. Tune settings on the following validation window.
3. Test on the next untouched period.
4. Move the window forward and repeat.
Keep the final test period sealed until model selection is complete. Test across different market conditions, including high-volatility periods, weak sector performance, and low-liquidity phases. Evaluate individual stocks and the portfolio rather than reporting only aggregate returns.
Track total return, annualized volatility, Sharpe and Sortino ratios, maximum drawdown, Calmar ratio, turnover, hit rate, average win/loss, exposure, and capacity. Include a benchmark and confidence intervals where possible. Stress costs by increasing slippage and reducing available liquidity. A strategy that survives conservative assumptions is more credible than one optimized for ideal fills.
Avoid common sources of false performance
The most frequent failures are methodological rather than algorithmic:
- Look-ahead bias: using future prices, revised classifications, or same-day information incorrectly.
- Survivorship bias: excluding failed or delisted companies.
- Overlapping data leakage: allowing normalization or feature engineering to use the full dataset.
- Overfitting: tuning indicators, rewards, and hyperparameters against the test period.
- Unrealistic execution: assuming every order fills at the close with no market impact.
- Ignoring corporate actions: mishandling splits, bonuses, dividends, or rights issues.
Use training-only scaling, point-in-time data, realistic order timing, and a transaction simulator. Review trades manually around earnings, exchange halts, and sharp gap moves. A version-controlled repository and a documented experiment log are essential; a guide to building a machine learning portfolio on GitHub offers a useful structure for presenting this work.
Paper trade before live deployment
Move from backtest to paper trading only after the strategy has passed out-of-sample tests. Paper trading should use the same data frequency, order rules, position limits, and cost model planned for production. Compare expected and realised prices, latency, rejected orders, partial fills, and data outages.
For live systems, separate research, signal generation, execution, monitoring, and emergency shutdown components. Log every state, action, order, fill, and model version. Start with small capital, impose daily loss limits, and require human approval for unusual trades. Never connect an untested agent directly to a broker account.
Practical 2026 checklist
Before claiming that the system works, confirm that you have:
- A precisely defined Gujarat chemical-stock universe.
- Point-in-time, adjusted data and a documented data dictionary.
- Indicators generated without leakage.
- A cost-aware simulator and explicit order timing.
- Simple benchmarks and a non-RL baseline.
- Walk-forward and unseen-period evaluation.
- Portfolio-level exposure and drawdown controls.
- Reproducible code, logs, seeds, and model versions.
- Paper-trading evidence before any live deployment.
The strongest project is not the one with the highest backtested return. It is the one that explains its assumptions, survives realistic tests, and fails safely when market conditions change.