Inverting a Trading Bot Reveals Flawed Signals
The article explores running an inverted version of a trading bot to diagnose whether it is extracting a genuine signal or a flawed one with the wrong sign. The results show the inverted bot outperforming the original, suggesting the original bot's signals were inverted.
Why it matters
This technique can help trading algorithm developers quickly identify the root cause of a losing strategy, distinguishing between lack of signal and flawed signal, which has important implications for model improvement.
Key Points
- 1Running an inverted version of a losing trading bot can reveal whether the original signals were flawed or genuinely lacking information
- 2The inverted bot showed a 70.59% win rate vs 15.79% for the original, indicating the original signals had the wrong sign
- 3The fee impact doubles when running parallel books, so the inverted book's gross profit needs to clear a higher bar
Details
The article describes setting up two parallel trading books - one executing the original bot signals, and one executing the exact opposite. After 24 hours, the inverted book was winning 70.59% of round-trips compared to 15.79% for the original. This asymmetric performance suggests the original bot was extracting a signal with the wrong sign, rather than lacking any signal at all. The authors note that the inverted book still has net losses due to the compounded fee impact of running two books, but the key metric is whether the inverted book's gross edge clears the higher fee floor by the 100-round-trip decision point. Running this inverted control experiment is presented as a low-cost way to diagnose the underlying issue with a losing bot - whether it's truly extracting no signal, or just extracting a flawed signal with the wrong sign.
No comments yet
Be the first to comment