CryptoPainter

There are already multiple projects asking me to test their AI trading frameworks. Let me just say a few points: 1. Present long-term real trading results; short-term meme coins without max drawdown don't mean anything, it's just survivor bias. Pick any coin that has already gone to zero and backtest it, the curve will look just as beautiful. 2. If the Sharpe ratio > 5, it can basically be determined that it is overfitting, look-ahead bias, or data leakage. Medallion usually only has 2-3, and if you’re getting 7 at home, you should be aware of that. 3. Testing with data from a crypto bull market is different from backtesting with data from major markets; it’s complete overfitting, and I generally don’t consider it. At the very least, it should work through the two bear markets of 2018 and 2022, and then run a walk-forward test to count as a strategy. 4. Transaction fees, slippage, and funding rates must be included. When you layer in Binance's maker/taker fees, VIP tiers, and BNB discounts, if the model is inaccurate, the backtest and real trading can differ by a factor of two in annualized returns, which is normal. 5. Strategy capacity is more worth watching than return rate. Just because it can handle $100,000 doesn’t mean it can handle $1,000,000. The depth of small coins is limited; when you enter the market, you can wipe out your own signals, which the backtest cannot reflect at all. 6. Indeed, crypto quant isn’t that competitive, but arbitrage opportunities are constantly being eroded—funding arbitrage, spot-futures basis, inter-exchange price differences have basically been cleaned out by market makers and HFTs. High-frequency trading is not feasible, and pure factors have no space left; the only paths left are trend and mean reversion. 7. Alpha has a half-life. If a strategy can still run after three months, it’s considered passing; if it lasts six months, it’s considered good; if it lasts a year, it’s likely just luck or your scale hasn’t reached a noticeable level. Don’t mistake the benefits of a single bull run for perpetual alpha; you’re not that impressive yet. 8. The "optimal parameters" from grid search are 99% overfitted. The truly stable parameters are those that can run regardless of the range you pick, not just those that work when fine-tuned to two decimal places. Parameter robustness is a hundred times more important than single-point returns; anyone who has done this understands. 9. The market structures of 2017 ICOs, 2020 DeFi summer, 2021 memes, 2022 LUNA/FTX, and 2023 AI narratives are completely different. The "patterns" you fitted in the previous phase can go to zero in a different regime, and you’ll even lose on fees. 10. Exchange risks are always bigger than you think. FTX going to zero, API rate limits, flash crashes, small exchanges running away, Binance suddenly delisting—these are all "one-time game-enders." You can’t withstand a single exchange disaster with a 50% annualized return; it doesn’t matter how great your strategy is, if you’re trading altcoins, you need to consider liquidity and "delisting risk." 11. The curves in backtests look beautiful, but when your account experiences three weeks of continuous value decline, 90% of people will shut down the program and manually adjust parameters. 12. Distinguish whether you are earning alpha or beta. In a bull market, everyone is a quant master; when a bear market comes, only those with beta get washed away. Isolate the long exposure to see the alpha curve; most so-called "strategies" have no alpha at all, they are just a disguised long on BTC with a bit of volatility. 13. There is a lot of false prosperity in ML within quant. LSTM, Transformers, and reinforcement learning are hyped to the sky, but in financial time series with extremely low SNR, a simple momentum factor with reasonable risk control can outperform your XGBoost that you’ve tuned a thousand times. Learning this is really tough; quant is truly a rare skill.

I just happened to come across this video, which gives a very concise explanation of why fish oil is a huge marketing scam... Saying that fish oil is effective is like saying that if you do a frog jump forward on your way home, it indeed shortens the distance, but what’s the point...