What makes a trading strategy good?
Part 1: The Answer
I analyzed over 100,000 different backtests.
No, that is not an exaggeration.
Over 5 years ago, I started building what would later become NexusTrade, a no-code algorithmic trading platform. Since my official launch on January 31st 2023, NexusTrade users have launched exactly 114,549 backtests.
Using natural language processing and data science techniques I analyzed the 100K+ backtests for patterns. The question I asked myself was simple.
What makes a trading strategy good? What makes a trading strategy bad?
After sifting through the noise, a clear pattern emerged. The difference between success and failure isn’t a secret indicator; it’s a fundamental approach to design. So, what’s the answer in a single sentence?
A good trading strategy executes a simple, logical, and complete plan tailored to a specific asset’s behavior, whereas a bad strategy is an over-complex, illogical, or incomplete set of rules that mistakes historical noise for a true market edge.
Let’s break that down. The data tells a compelling story, best understood by looking at what doesn’t work first.
Full Disclosure: I’m the founder of NexusTrade, the platform that generated this data. This analysis uses real user strategies from our production database. All code and methodology are open-source.
The Hallmarks of a Bad Strategy
The worst-performing strategies weren’t unlucky; they were fundamentally flawed. They consistently shared these traits:
- Illogical or Impossible Rules: Many failing strategies contained conditions that could never be met, like Buy “when 0 < 0”, or conditions that were always true, like “Sell always”. These are not strategies; they are programming errors that guarantee failure through either inaction or excessive trading costs.
- The “Bag Holder” Problem: A huge number of bad strategies had rules to buy but no clear rules to sell. A “buy the dip” strategy without a profit-taking or stop-loss rule is just a plan to accumulate a depreciating asset. A strategy must be a complete plan for a trade’s entire lifecycle.
- Over-Complexity and Contradiction: The most convoluted strategies were often the worst. One strategy had 13 different rules mixing dollar-cost averaging, percentage-based buys, and conflicting moving averages. This “kitchen sink” approach is a classic sign of overfitting to historical noise and is almost guaranteed to fail in live markets.
The Anatomy of a Good Strategy
In stark contrast, the elite and high sharpe strategies were built on clear, logical, and often simple principles that were context-aware.
- Context is King: Momentum vs. Mean Reversion: The data showed a powerful pattern: This was the 2025 classification result. I withdrew the stock-versus-index recommendation in September 2026 because the result changed when the classifier's definition changed. The examples below illustrate rule types; they do not establish which style works best for either asset group.
- For example, a top condition was GOOG Price > 20 Day GOOG Bollinger Band and 15 Day GOOG RSI > 70 — a classic momentum breakout signal on a specific stock. Conversely, another top condition was 30 Day SPY Rate of Change ≤ -5 — a classic “buy the dip” signal on the entire market. Good strategies don’t apply one logic to everything; they match the logic to the asset’s behavior.
- Simplicity and a Clear Thesis: The best conditions were easy to understand. They represented a clear market thesis, like “buy when the market is fearful” or “join a strong trend.” This simplicity makes them more robust and less likely to be overfitted.
- Intelligent Risk Management: Some of the smartest conditions weren’t just about buying dips; they were about buying dips in an established uptrend. An elite-performing condition was SPY Price > 200 Day SPY SMA and 30 Day SPY Rate of Change ≤ -5. This logic is brilliant in its simplicity: it waits for the market (SPY) to pull back significantly (≤ -5%), but will only consider this a buying opportunity if the overall market is still in a long-term uptrend (above its 200-day moving average). This rule systematically avoids catching a falling knife at the start of a new bear market, dramatically improving risk-adjusted returns by filtering out the most dangerous drawdowns.
My Methodology: From 100K Backtests to Actionable Insights
Data Pipeline:
- Population: 114,549 backtests run between Jan 31st, 2023 — Nov 3rd, 2025
- Tested Against: ~20 years of historical market data (primarily 2004–2025). Includes: 2008 Financial Crisis, 2010s Bull Market, 2020 COVID Crash, 2021–2022 Bear Market, 2023–2024 Recovery
- Normalization: Intelligent number rounding to group similar strategies (e.g., “RSI > 69.8” and “RSI > 70.1” → “RSI > 70”)
- Metrics: Annualized returns, Sharpe ratio, Calmar ratio (risk-adjusted)
- Diversity Selection: TF-IDF + cosine similarity to prevent showing 100 variations of the same RSI strategy
- Corpuses: Top 1% (elite), Top 20% (high performers), Bottom 10% (terrible)
Statistical Rigor:
- Used median (not mean) to avoid outlier bias.
- Filtered infinite/invalid values (e.g., Calmar > 1000).
- Historical inclusion floor: at least five runs per component. An occurrence threshold alone does not establish statistical validity, especially when runs include cloned strategies or selected favorable periods.
How Confident Am I?
The patterns are statistically robust:
- Elite strategies (top 1%): Median Sharpe 3.2 vs. 0.8 overall (4x better)
- Terrible strategies (bottom 10%): Median Sharpe -0.4 (consistently lose money)
- Sample size: Elite conclusions based on 1,145+ unique strategy components
- Consistency: 73% of “elite” components appear in multiple high-Sharpe backtests
Translation: These aren’t flukes. The patterns are real.
What This Analysis Doesn’t Tell You
Be skeptical. This analysis has limitations:
- Survivorship Bias: Only analyzed completed backtests (users may delete failures)
- Data Quality: Pre-2004 market data is limited/lower quality in our database
- Recency Bias: Most users test recent 5–10 year periods, not the full 20 years available
- Market Evolution: Strategies that worked 2004–2008 may not work in modern algorithmic markets
- Execution Gap: Backtest results ≠ live trading (slippage, fees, psychological factors)
Bottom line: These patterns are strong directional indicators, not guarantees.
Conclusion: Your Three Takeaways
After analyzing over 100,000 backtests, the conclusion is clear: the secret to a good trading strategy isn’t a magic indicator, but a disciplined approach.
- Simplicity Over Complexity: Simple rules are easier to inspect and explain. The September 2026 corpus did not show a monotonic performance decline as rule count rose: its median Sharpe was 0.932 for one rule, 0.606 for two rules and 0.927 for five to eight. Those are observational groups, so rule count alone is not a performance guarantee.
- Context is Non-Negotiable: The best strategies are tailored to the specific behavior of the asset being traded. Don’t apply a stock strategy to a market index and expect it to work.
- A Strategy Must Be a Complete Plan: Every entry rule must be paired with a well-defined exit plan for both taking profits and cutting losses.
Take Action
Want to test these principles? Three options:
- DIY: Use the analysis code (Python/Polars) with your own backtest data
- Study Further: Research academic papers on momentum vs mean-reversion strategies
- Quick Start: Try the no-code platform that generated this dataset (NexusTrade)
All analysis code is open-source and available on GitHub.
The market data for this analysis comes from EODHD. Need high-quality data for your investment strategies? Try it out for free today!
No comments yet.