More than one project has approached me to test their AI trading architecture Let me briefly say a few points 1. Take a long-term stock, short-term monster coins without max dd do not represent anything, they are just survivorship bias. Pick any coin that has already been reset to zero and test it upside down. The curve is as beautiful as ever If Sharpe ratio>5, it can basically be determined to be overfitting, look ahead bias, or data leakage. Medallion is only 2-3 all year round, and if you run 7 alone at home, you need to have a b-number in your mind 3. Using a piece of data from a crypto bull market for testing is different from using a piece of data from a major market for backtesting. It's completely overfitting and I don't really look at it. At least two bear markets in 2018 and 2022 can work together, and running another walk forward is considered a strategy 4. Transaction fees, slippage, and funding rates all need to be included. Binance's maker/taker, VIP tier, and BNB discount are calculated layer by layer. If the model is inaccurate, the difference between backtesting and actual trading can be doubled annualized, which is the norm 5. Strategy capacity is more worth considering than yield. Running with 100000 dollars doesn't mean running with 1 million dollars. The depth of small coins is just a bit, as soon as you enter the market, you will destroy your own signal, and there will be no reflection in the backtest 6. It is true that crypto quant is not as popular, but arbitrage opportunities have been constantly eroded - funding arb, spot futures basis, and cross exchange spreads have been basically eaten up by market makers and HFT. High frequency cannot be achieved, and there is no room for pure factors. The only things that can be done are the old paths of trend and mean reversal Alpha has a half-life. The strategy can still run after three months of being launched, which is considered passing; Half a year still, pretty good; There is still one year left, most likely due to luck or your scale not yet reaching the level of attention. Don't treat a bull run dividend as a perpetual alpha, you're not that awesome yet 99% of the "optimal parameters" obtained by Grid search are overfitting. The truly stable parameters are those that can run freely within a range, rather than working to two decimal places. Parameter robustness is 100 times more important than single point returns, and those who have done this understand 9. 2017 ICO, 2020 DeFi summer, 2021 meme, 2022 LUNA/FTX, 2023 AI narrative, each segment has a completely different market structure. The 'pattern' you fitted in the previous paragraph can be reset to zero by changing the regime, resulting in a loss of transaction fees 10. Exchange risks are always greater than you think. FTX zeroing, API flow limiting, pin bursting, small office running away, and sudden removal of Binance are all things that end the game in one go. Your annualized 50% cannot withstand an exchange crash, which has nothing to do with your strong strategy. Those who do knockoffs need to consider liquidity and "delisting risk" 11. The curve fluctuation on backtest looks beautiful, but if the net asset value drops for three consecutive weeks in one's own account, 90% of people will close the program and manually adjust the parameters 12. Distinguish whether you are earning alpha or beta. In a bull market, everyone is a quant master, but in a bear market, those who are only beta are washed away. Stripping out the exposure of long positions and looking at the alpha curve separately, most so-called "strategies" do not have alpha at all, which is essentially adding a little bit of volatility to long BTC. 13. ML has a lot of false prosperity in Quant. LSTM、Transformer、 Reinforcement learning is being hyped up, but in financial time series with extremely low SNR, a simple momentum factor combined with reasonable risk control can surpass XGBoost that you have adjusted 10000 times When it comes to learning JB, Quant is truly a rare achievement