Each week, companies issue commercial paper with maturities from 1 day to 270 days. For four issuer types we track the mix, meaning the share of dollars in each of six maturity buckets. The question is whether next week's mix can be forecast, and whether optimal transport (OT) methods add anything.
We place the buckets on a log-days axis and compare two mixes by the squared 2-Wasserstein distance. That distance splits exactly into location, scale and shape parts:
W₂² = (m₁−m₂)² + (s₁−s₂)² + 2·s₁·s₂·(1−ρ)| check | result |
|---|---|
| Week-to-week change vs. sampling noise | 4.0× larger: real movement |
| Share of change that is shape | 67%: not just a shift |
| Power to detect a 5% gain over 927 weeks | 0.50 (just at the bar); 10% gain: 0.98 |
| Smoothing (EWMA) vs. long-run average | 58% lower loss: forecastable |
Caveat: splitting each week by days gives a noise level only 22% below the weekly change, so a lot of the movement is day-level noise. Expect modest gains.
Nothing after 2008 has been scored. The out-of-sample test runs once, and only after approval.