CorrAI

Executive Summary
In this article, I explore a $BTC trading strategy using Z-score normalization—a well-established tool in mean-reversion analysis. I built and tested this strategy on a no-code platform called CorrAI, and currently, forward testing it. While the backtest returns and metrics like Sharpe ratio (3.47) and Calmar ratio (16.94) are compelling, a closer look at the distribution of returns reveals possible overfitting and risk concentration in outliers. The following breakdown is not an endorsement of the strategy but a case study in statistical due diligence.
Strategy Design
Conceptual Framework
Z-score normalization rescales time series data by subtracting the mean and dividing by the standard deviation:
Z = (x - μ) / σ
Where:
x = observed price
μ = rolling mean
σ = rolling standard deviation

It’s a common technique for mean-reversion strategies, highlighting deviations from historical norms.
Strategy Formula (No-Code Expression)
Using a no-code environment, I translated the Z-score into a form that avoids parentheses:
1h | btc | close / 1h | btc | close # STDDEV 120 1 - 1h | btc | close # LINEARREG 120 / 1h | btc | close # STDDEV 120 1
Trading Rules
Stop Loss: 5%
Trailing Stop Loss: 1%
Entry: Z-score > 0
Exit: Z-score <= -1.5

Backtest Overview
Period: Aug 5, 2024 – May 14, 2025 (283 Days)
- Total Return: 196.94%
- CAGR: 308.97%
- Sharpe Ratio: 3.47
- Calmar Ratio: 16.94
- Sortino Ratio: 4.05
- Max Drawdown: -18.24%
- Time in Market: 77.1%
While the equity curve appears consistent, deeper trade-level diagnostics are necessary.

Risk & Trade-Level Metrics
- Total Trades: 391
- Win Rate: 43.73%
- Profit Factor: 1.46
- Average Return per Trade: 0.27%
- Average Holding Time: ~13.3 hours
- Max Losing Streak: 8
Despite promising performance ratios, a low win rate and short holding time hint at risk concentration.
PnL Distribution Analysis
- Mean Return: 0.30%
- Median Return: -0.24%
- Around 75% of trades are losing or near-zero
- Profits come from rare outliers (long right-tail events)

A smooth equity curve doesn’t always imply signal. In this case, profitability depends heavily on irregular, high-gain events—suggesting fragility and potential overfitting.
Monthly Performance Snapshot
Month | Strategy Return | Buy & Hold | Delta
Jan | 17.9% | 9.1% | +8.9%
Feb | 19.1% | -17.1% | +36.2%
Mar | -0.5% | -1.5% | +1.0%
Apr | 12.2% | 12.9% | -0.7%
May | 5.5% | 10.0% | -4.5%

Interpretation
Pros:
- Straightforward implementation
- High-level metrics look appealing
- Useful as a sandbox for learning factor testing
Cons:
- High dependency on rare winners
- Trade distribution skewed toward loss
- No multi-factor validation
Takeaway: surface-level metrics can obscure fragile foundations. Always check the return distribution.
Next Steps & Discussion Points
Some ways to build upon this analysis:
- Normalize non-price data (on-chain wallet metrics, volume)
- Add volatility filters or trend classifiers
- Validate over multiple assets or timeframes
- Perform walk-forward analysis to test real-world resilience
Curious to hear how others might reduce reliance on tail events or if you've explored similar setups using Z-score normalization.