Why My AI Trading Strategy, Which Returned 23% in 2 Months in a Simulated Environment, is Promising

Backtesting and simulated trading are essential steps in evaluating a trading strategy. However, not all backtests are created equal—many strategies exaggerate their profitability by making unrealistic assumptions. My AI-based trading strategy, which yielded a 23% return over two months, has been tested with transparency and fairness, making it a promising approach despite being in a simulated environment.

Unlike many retail traders, I take a hedge fund-like approach, combining fundamental analysis, sentiment analysis, and technical indicators to develop a structured and data-driven trading strategy. This multi-layered methodology increases the robustness of the system, ensuring that trading decisions are based on objective signals rather than speculation.

In this article, I’ll outline the rigorous steps I have taken to ensure accuracy in backtesting, discuss the results, and highlight challenges that might arise when transitioning to live trading.

Steps Taken to Ensure a Fair and Transparent Backtest

To ensure that my simulated trading results are as realistic as possible, I incorporated several key measures:

1. Conservative Entry and Exit Price Assumptions

One of the most common mistakes in backtesting is assuming perfect trade execution at ideal prices. To avoid this:

• I assumed entry at the highest price in the first minute after a relevant event rather than a mid-point or an average price. This prevents unrealistic assumptions about entering trades at the best possible moment.

• If neither the stop-loss nor the target price was triggered, I exited at the closing price of the second trading day to account for price movements stabilizing over time.

2. Custom Python Scripts to Ensure Objectivity

Since many paper trading platforms do not support micro-cap stocks—where my AI strategy identified many opportunities—I built a custom Python-based backtesting script that processed minute-wise data. This script:

• Ensured strict rule-based execution without manual intervention or bias.

• Simulated realistic price movements and execution conditions.

• Applied stop-loss and target logic automatically based on predefined parameters.

3. Exclusion of Trades Triggered by After-Hours Events

Institutional investors move markets significantly based on after-hours developments. By the time retail traders can react at market open, a stock’s price may already reflect the impact. To avoid unrealistic entries:

• I excluded all trades triggered by events that occurred outside regular trading hours.

• This ensures that my assumed entry prices reflect a scenario where a retail trader could actually execute the trade.

4. Implementation of a Trailing Stop-Loss

Trailing stop-losses are essential for risk management but are often ignored in unrealistic backtests. I incorporated a dynamic trailing stop mechanism, which:

• Allowed profitable trades to lock in gains while minimizing downside risk.

• Increased the number of break-even trades, reflecting a realistic outcome where stops are adjusted as the stock moves favorably.

5. No Hindsight Bias or Curve Fitting

Many backtests become overly optimized to historical data, making them unreliable for live trading. My approach:

• Used a predefined rule set before running the backtest—no parameters were adjusted to improve past results retroactively.

• Ensured every trade was executed based on real-time AI-generated signals, rather than selecting trades after seeing price movements.

Results of the AI-Based Trading Strategy

Across 96 trades over two months, my strategy achieved:

• Win Rate: 44.79% (percentage of trades that were profitable)

• Loss Rate: 34.38% (percentage of losing trades)

• Break-Even Rate: 20.83% (percentage of trades that exited at break-even due to the trailing stop-loss)

• Win-Loss Ratio: 1.30 (for every 1 losing trade, I had 1.30 winning trades)

• Total Return: 23%

• Expected Value: 0.34 ( EV = 0.3437 means that for every ₹1 I risk per trade, I can expect an average return of ₹0.34)

• Max Drawdown (MDD)- 7.84% meaning that during the worst trading phase of the simulation the largest peak-to-trough decline was 7.84% before recovery

• With a 1:2 risk-reward ratio, my average loss is 1% per losing trade so if we assume a fixed-risk allocation model, where i risk a constant percentage of capital per trade, then:

• Risk per trade ≈ 1% (as per my setup)

• This means for each losing trade, I lost 1% of the portfolio

This performance is encouraging because it reflects :

• Realistic win rate rather than an artificially inflated success rate. Unlike deceptive backtests that report nearly 100% win rates, my results acknowledge the presence of losses and break-even trades, which are inevitable in real-world trading.

• Since EV is positive, this suggests that over a large number of trades, my strategy is profitable.

• MDD of 7.84% is reasonable given the short-term nature of this strategy and aligns with typical risk management principles.

• Risk per trade is 1%, meaning controlled exposure per trade, which aligns well with professional risk management.

Additionally, my risk-reward ratio has been carefully structured to ensure that profitable trades, on average, significantly outweigh losing trades. This means that even with a moderate win rate, the overall strategy remains profitable over time.

Challenges in Transitioning to Live Trading

While my AI-based strategy has demonstrated promise in backtesting, real-world trading presents additional challenges that must be addressed:

1. Liquidity & Slippage Risks

• My strategy often identified small and micro-cap stocks, which can have limited liquidity.

• In live trading, executing orders at the expected price may not always be possible, leading to slippage that can reduce profitability.

2. Market Reaction Speed

• Institutional traders and high-frequency algorithms react to news in milliseconds, whereas retail traders may experience slight execution delays.

• My assumption of entering at the first-minute high is a conservative estimate, but real-world execution might still lead to less favorable entry points.

3. Market Impact Beyond Two Trading Days

• While I assumed that price impact stabilizes within two trading days, some events may lead to extended volatility, requiring adjustments to my exit criteria in a live setup.

4. Transaction Costs & Brokerage Fees

• The backtest did not account for brokerage commissions or transaction fees, which could slightly lower the net returns in real trading.

Final Thoughts: Why These Results Are Encouraging

Unlike many misleading trading strategies that claim unrealistic success, my approach has been:

• Transparent—Clearly disclosing assumptions and limitations.

• Realistic—Accounting for execution delays, conservative pricing, and real-world risks.

• Systematic—Based on AI-driven sentiment analysis, without hindsight bias or post-hoc adjustments.

A disclaimer like this ensures expectations remain realistic: "Results are based on simulated execution using conservative entry assumptions (first-minute high after news). Actual trading conditions, including slippage and liquidity constraints, may impact real-world performance."

While live implementation will require additional refinements, this AI-driven strategy shows strong promise as a viable trading approach. The key takeaway is that honest, well-tested strategies are more valuable than unrealistic claims of guaranteed success.

To know more about my trading strategy , read my blog - Leveraging AI/ML for Finance and Trading: A Journey from ML Models to a 23% Gain in simulated environment with an AI-Powered Strategy..