I have been backtesting 24 different combinations of stats to bet on 5 years worth of historical lines. One of my offense models has a 6% return. My best pitching model has a -0.5% return. When I do different ratios like 30% hitting 70% pitching or whatever, I get even better results than either by themselves.

The thing is, I can't quite understand why that would be the case. Doesn't that seem counterintuitive? Has anyone had a similar experience?

Not enough info. What are your models producing, prop lines? When you talk about mixing them, what does that mean?

Straight ML bets. And combining a hitting + pitching stat gives a higher win rate than either one alone.

I would still need more details to be able to explain what's happening. Assume that we're laymen when it comes to sports modeling and make it so that we can use only our knowledge of probability to give an answer.

What I can tell you is that incorporating something that's truly -EV or noise doesn't improve your overall EV.

However, if you're adjusting some kind of weight assigned to each stat, then that's a different strategy altogether and wouldn't necessarily be related to the EV of a strategy that uses weights of 0 or 100 for each stat.

Or if your 70% means "70% of bets produced by the pitching stat" then it's certainly possible that the 30% of the discarded bets happened to include a disproportionate amount of losing bets by chance alone.

Ok, so my combination of offensive stats that I use returns 6% just factoring offense.
The defense stats, returns -.5% using only defense

Using a 70% weight of the offensive stat + 30% of the defensive stat gives a better number than both separately.

It's still betting the same number of games.

Ok so then if I understand you correctly, nothing weird would have to be going on for this to be a real effect. As an extreme example, if I try to predict the winner based only on who has the better starting pitcher, that will be a -EV model because that's not the only factor. But that doesn't mean that factoring the starting pitcher is a blanket -EV thing to do or that it's not a real factor. It just means that I gave that factor the wrong weight relative to other factors.

Theoretical question:
Imagine that for a certain sport one used their domain expertise to isolate 3 stats (A, B, and C) they think will have strong predictive value. They perform a linear regression to get the ideal weighting of each stat, and when backtesting wagers of that 3-feature model it has a positive return on investment. However, when they look back at the regression report they notice that stat C has a p-value greater than 0.05 so they remove it as a feature from the model, but then the wagering backtest has a slightly lower ROI, say around 0.5 to 1% less.

Would you assign more importance to the information provided by the p-value and continue to leave stat C out of the model, or would you assign more importance to the backtest ROI and leave stat C in the model?