I took my v9 LSTM into a live trading competition

People had asked me whether my stock prediction model really worked. I wanted to see what would happen if I took it out of the notebook and used it to make a portfolio decision in the ACM ICAIF 2026 Trading Agent Competition.

I used my v9 LSTM to rank 30 US stocks and turn those forecasts into portfolio weights. The opening allocation executed on October 12. I packaged the code, trained models and inputs so someone else could replay the decision offline. The replay reproduced the forecasts and portfolio weights exactly. Source code and reproduction guide

How the portfolio behaved

The account started with $1 million on October 12 and closed at $994,683.22 on October 30, a loss of $5,316.78 (0.53%) after fees. At its highest recorded intraday valuation, it was up $14,072. The largest peak-to-trough decline was 2.20%. Competition results

I had planned to rebalance daily. I submitted the opening allocation, which bought 30 stocks and incurred $989.02 in fees. The portfolio carried those holdings forward through the remaining 104 rounds. It ended October 29 up $9,916, then lost $15,233 on the last day. With the positions held throughout, that last-day move came from changes in their market value.

The individual holdings moved in very different directions. Eli Lilly (LLY) added $9,187, while Intel (INTC) reduced the account value by $16,594. The chart shows each position’s contribution at the final close, including its share of opening fees.

Published metric	Result
Cumulative return after fees	−0.532%
Sharpe ratio	−0.857
Maximum drawdown	2.20%
Average turnover per round	0.942%

Codabench published the four metrics. My entry received no overall rank under the final scoring rule, which required a positive cumulative return. Published leaderboard data

For my next experiment, I want to test the daily decision and upload workflow end to end and verify each execution. That would let me measure the strategy I originally planned to run.