skaters · non-price benchmark

How robust is the win? Slice it and see.

Every non-price FRED change-series scored so far, one-step-ahead, held-out. Pick any subset — sample it at random, balance by category, filter by keyword, seasonality, tails, whatever — and the per-series win-rate of laplace against each opponent recomputes live, with a bootstrap 90% band. A robust edge stays high and tight no matter how you cut the data.

Series selected

laplace mean logpdf

–

Opponents beaten (LL >50%)

–

Accuracy vs. speed frontier

mean held-out log-likelihood over the selected subset · speed measured, fixed

Robustness vs.

laplace win-rate per bin · vs

Head-to-head win-rate

per-series, laplace vs each opponent

metric laplace wins >50% <50% │ whisker = bootstrap 90% CI · vertical line = 50%

Opponent	laplace win-rate (per series)	rate	N