Why Retraining AI Trading Models Matters

Evidence from Our ES & NQ Strategies

It’s intuitive to believe that retraining AI trading models matters. Markets evolve, and so should our signals but it’s always better to find the data that validates or disputes the assumptions. 

The charts below quantify this effect for our ES (S&P 500 Futures) and NQ (Nasdaq Futures) strategies, comparing Sharpe Ratios between retrained and non-retrained models over the past four quarters.

• You can find more details about our AI trading model, Pythia, in our recent blog post.
• Details about the trading strategy used to calculate the Sharpe ratios and additional performance metrics can be found in our recent performance benchmark.

What We See

• Retraining consistently improves performance.
  Across both markets, the retrained models maintain higher Sharpe Ratios throughout the period, even through challenging quarters.
  
• Q1 and Q2 were especially volatile.
  These quarters put both models to the test, with non-retrained versions dropping sharply (even negative for ES and NQ). Retraining helped both recover and stabilize faster.
  
• Differences between ES and NQ.
  The ES baseline model shows a strong rebound in Q3 achieving a slightly better Sharpe Ratio (1.518) to the Sharpe Ratio for 24Q4 (1.253). This is still behind the retrained signal (3.7) but it shows stronger resiliency of the baseline model. NQ’s recovery is more moderate. This divergence highlights that retraining impact can be asset-specific — reflecting the unique microstructure and sentiment dynamics of each market.

What’s Next

• Retraining frequency optimization: what happens if we retrain monthly, bi-weekly, or even daily?
  
• Extended backtesting: understanding how retraining cadence interacts with different market regimes over longer time horizons.
  

The goal is to find the optimal balance between adaptability and stability, ensuring our models remain responsive to changing conditions without overfitting to short-term noise.

Takeaway

Retraining isn’t just a technical maintenance task. It’s a core part of model lifecycle management in live trading systems, directly impacting alpha preservation, drawdown control, and risk-adjusted performance.