Back

Kelly Sizing Degenerates to Equal Weight - What should I do?

Building a systematic long-only S&P 500 stock ranker and running into a Kelly sizing problem I'd love input on.

The pipeline:  
- Universe: ~479 S&P 500 stocks scored monthly
- Signal: 62.5% quality composite (ROIC, FCF margin, gross margin, accruals) + 37.5% XGBoost classifier (1m horizon), both OOS IC-positive (Quality IR=0.60, XGBoost  IR=0.36)
- Sizing: half-Kelly per position, capped at 15%, then Ledoit-Wolf vol constraint (target 15% ann.), sector cap 30%, min position 2%

The problem:
With typical model outputs (p_win ≈ 0.55–0.70, avg_win ≈ 2–4%, avg_loss ≈ -2–3%), the half-Kelly formula almost always yields >15% per position. Every stock in the top-20  hits the cap. After portfolio normalization (long-only, no leverage), all positions end up at ~5% each, effectively equal weight, with Kelly adding zero differentiation.

The two options I'm considering:
1. Score-proportional sizing — ditch Kelly, size each position proportional to its "ensemble_score" rank (score-weighted allocation), still subject to the 15% individual cap  and vol constraint
2. Lower the cap — reduce "MAX_POSITION" from 15% to ~5–7%, run 10–12 positions, let Kelly differentiate within a tighter band

My concern with option 1: Abandoning Kelly feels like throwing away the probabilistic interpretation of the classifier output (p_win, avg_win, avg_loss) which took real  work to calibrate.

My concern with option 2: With 10–12 positions and a 5% cap, even the best Kelly estimate barely moves the needle, the constraint is still doing most of the work.

Is there a smarter way to scale Kelly outputs that actually survive a max-position cap? Or is score-proportional sizing the pragmatic answer for this type of factor model?