Hello,
I hope you’re all doing well!
Like the title says, I made a new trading program. This took me a while (over a month of nonstop work).
My aim was to find structural alpha, not just for one symbol. I wanted something more ambitious: a structural alpha
that could apply to any crypto asset—and I think I made it. I believe this alpha is related to leverage and
liquidation cascades (not 100% sure). One hope I have is that this could help increase market liquidity and reduce
those events if that’s actually the reason.
Right now I have one model I trust, for FARTCOIN. The reason is I started with Bybit data, and Bybit leads this coin.
I’m currently downloading Binance data so I can train models for coins where Binance leads. If you train models for
other coins, you’ll notice some strategy parameters are very similar—this is exactly what you want when you’re trading
structural, market‑wide alpha.
The program uses gradient boosting with DART (worked best for order book features). It analyzes several features with
normalized TA only for context, so the model is price‑agnostic. I built microstructure and L2 order book features,
plus broader context features to help it navigate different market regimes (inventory + context). The focus was
robustness—something that can survive regime shifts under extreme scenarios.
Right now the model performs well. I’ve deployed live strategies with real money and the results match the backtests.
That said, I haven’t thoroughly tested live execution for the latest models (I added more complexity, like multiple
positions per side). Because of that, I strongly recommend running in --dry-mode first (if you use the Python
scripts), then --testnet. If you use the launcher, just run testnet (there’s an option for that). When testing live
execution, monitor entries and make sure everything works smoothly when positions open.
If you have suggestions or just want to chat, feel free to reach out:
https://github.com/wizrdfab/PurpleSky I added contact info there and I’d be happy to help. Maybe we can even form a small community around this.
If you like the program or find it useful and want to support or contribute, I added donation info in the GitHub
README. You can also register for Bybit with my referral link:
https://www.bybit.com/invite?ref=14VP14ZLink for the release + source code:
https://github.com/wizrdfab/PurpleSky/releases/tag/v0.1.0-alphaThanks!
