Key insights
- The author discusses the lack of standards in crypto backtesting, highlighting issues like omission of fees and cherry-picking start dates. They share their own AI-driven crypto index backtest results, revealing a high drawdown and low Sharpe ratio, emphasizing the importance of considering fees. While the backtest outperforms a BTC/ETH benchmark, the author cautions against over-interpreting the results due to limited live data, suggesting a bearish signal for similar crypto strategies.
If you've been here for some time, you know how it goes: Someone shares their equity graph claiming that their method is “10x since 2020,” the graph is nice and clear, and somehow somewhere deep down in the thread you’ll realize that they haven’t taken into account fees or started their backtest when the markets were rising. This is not always done with malicious intent, sometimes it’s simply that people are uncertain about what to share, or feel awkward about the ugly stats. I constructed an AI-driven crypto index - machine learning predictions, risk-parity weighting optimization, monthly rebalancing, weight publication with run IDs to make results reproducible. Backtesting spans August 2021 to December 2025, while the live data starts in January 2026. Here are my less-than-stellar stats: 73.7% maximum drawdown, 0.22 Sharpe, and fees have been excluded from the backtest.
That last part matters more than people acknowledge. Monthly rebalancing across 10 crypto assets means crossing spreads 10 times per cycle. Depending on the assets and the venue, that's easily 0.5-1.5% in friction per rebalance. Over 4.5 years of backtest, that compounds into a meaningful gap between what the model shows and what you'd have actually made.
The 120.3% total return vs 14.2% for a BTC+ETH 50/50 benchmark looks good. But I have three months of live data. Three months. That's not a validation, that's a start.
My intention is not to promote the project, but rather that I feel there is an embarrassing lack of standards when it comes to an "honest backtest" of any algorithmic trading strategy in cryptocurrency, and building something makes you realize just how easily manipulated the results can be.