|
Let me first give a little background about me. I am a software developer by profession.My background in quant is limited (1 3 month internship at a firm during college).I started quant trading as a side project mostly cryptos.
In the last year and a half i have found > 40 signals on bitcoin/eth at 30 min ; 1 hr horizon with sharpe between 1 and 2. However none of them performed well live/paper trading. The strategies were based on technical indicators , momentum , reversion , volatility filters etc. mostly sourced from journals and intuition.
Earlier i used to run the backtest on whole dataset, but after reading about overfitting i started using validation sets (3 months ),even multiple sets. However strategies filtered through these sets (> 20) have also sucked in live . They are either flat or start going down rapidly in live/paper trading. I had hoped that atleast some of them will perform live but none are.
Is there something systematically wrong about my approach. The strategies arent highly correlated either and are based on different ideas.
|
|
|
|
|
Well, what is the source of the poor performance? Is it losing money from slippage and fees? Or are the trades just flat out wrong?
Assuming it's the latter, those signals/indicators are very backwards looking. In other words, when you look at something like bollinger bands, it's easy to say -- Wow, if I bought/short every time it went outside of the band, it would've made a killing.
But the problem is, with bollinger bands specifically, those signals only can be shown AFTER the fact. So the 2 std dev move that looked like a signal, only became a 2 std dev move after all the other future points had been calculated. For example, many times when trying this strategy irl, a data point will break the lower band creating a buy signal, but then the price just keeps falling-- then 30 minutes later, the same price you entered the trade in is still within the bands, even though it was below the band when you first entered it.
The overwhelming majority of indicators like the ones you described fall victim to that bias. Your strategy has to have actual economic reasoning. At the end of the day, the prices may be easy to just think of as just pattern-bound numbers on a screen, but they always move for legitimate economic reasons. Many engineers dive straight into trading without knowing much market theory.
Here's what I recommend. Since you already can automate trades, go after trades that have actual economic sense -- dual class arbitrage for example. This arbitrage captures the spread in returns of two stocks that have dual-listings. It hasn't been arbed away yet and you can get data for it from here: Quantitative Global
Here are other posts that go into its theory/implementation:
- Theory + Manual Implementation
- Theory + Python Implementation |
|
|
|
The source of poor performance is wrong trades.
Your point about economic interpretation is correct but most of the things i have tried and read about are price action based strategies and i could not think of much economic logic in bitcoin/eth at intraday level.
I have developed price action based strategies before on fx which though were weak did work on live also. I am not sure why cryptos are behaving this way. Is the high volatility masking the noise as backtests signals.
Could you suggest something to atleast partially mitigate this. |
|
|
|
