How should you map model confidence to position size?

A safer sizing rule treats model confidence as one input to size, not as permission to lever up without limit. The score should first be calibrated, then translated into expected value or rank strength, and only then capped by volatility, liquidity, portfolio overlap, and drawdown tolerance.

Models are often overconfident in exactly the tails where naive sizing rules get most aggressive. If you map a raw score directly to exposure, you can accidentally concentrate risk in the part of the forecast that has the least reliable sample support.

Useful mappings are usually monotonic, capped, and boring. Teams often size by calibrated buckets, by expected edge divided by risk, or by a rank-to-weight curve that compresses the tails instead of letting them dominate the book.

Check how the equity path changes if the high-confidence bucket is slightly less reliable than expected, if execution is slower than the research assumed, or if several sleeves peak together. A sizing rule that only works when the confidence map is perfect is already too fragile.