Does the projected J-curve actually look like history?
A model can show a real pricing edge and still project the wrong shape — the wrong timing of calls and distributions, the wrong trough, the wrong overall spread of outcomes. We scored all five pricing models on four separate shape- and distribution-fidelity checks, on the same held-out funds used elsewhere on this platform, and no model wins on all four.
Four different questions, not one
Each fund's projected net-cashflow path (calls, distributions, and terminal NAV, combined into one signed series and scaled by the fund's NAV at the pricing date) was compared against that same fund's own realized future path, held out from calibration exactly as every other accuracy number on this platform is. Four metrics, each asking a genuinely different question — see the full methodology paper for the formulas:
- DTW (Dynamic Time Warping) — cumulative shape distance after the best nonlinear time-alignment. Lower is better. The closest thing here to a direct J-curve-timing test.
- Fréchet distance — the single worst pointwise gap along the best alignment. Lower is better. Catches one badly mistimed event that a cumulative measure like DTW can average away.
- Lag-optimal cross-correlation — directional tracking, allowing a small uniform time-shift. Higher is better. We also report the share of funds needing zero shift to best align — a read on whether a model's own pacing assumption is usually already right.
- Wasserstein distance — not a path comparison at all: the distance between the distribution of a model's own predicted totals across all held-out funds and the distribution of what those funds actually did. Lower is better. This is the population-level calibration check — where a systematic over- or under-projection shows up even when per-fund path shapes look fine.
The results, all five models
150 held-out mature funds (the same 2009-2016 vintage test population used throughout this platform's accuracy reporting), no look-ahead into the future.
| Model | DTW | Fréchet | Cross-corr. | Zero-lag share | Wasserstein |
|---|---|---|---|---|---|
| TA | 0.0328 | 0.3897 | 0.477 | 34.7% | 0.7162 |
| Equisect Bayesian | 0.0335 | 0.4168 | 0.489 | 33.3% | 0.8869 |
| Equisect Cohort | 0.0429 | 0.3668 | 0.568 | 66.0% | 0.5088 |
| Equisect Ledger | 0.0446 | 0.3467 | 0.448 | 56.4% | 0.3602 |
| Equisect Dynamic | 0.0305 | 0.3714 | 0.571 | 53.3% | 1.0052 |
Bold reading: Equisect Dynamic has the best cumulative shape-tracking (DTW) and the best directional correlation of all five — genuinely good path fidelity — and by far the worst population-level calibration (Wasserstein), more than double the next-worst model. Equisect Cohort is the most balanced across all four — strongest zero-lag share, mid-to-good everywhere else. Equisect Ledger has the best worst-case robustness (Fréchet) and the best population calibration (Wasserstein), but the weakest cumulative and directional path-tracking of the five. TA and Equisect Bayesian are unremarkable on every metric here — no standout strength or weakness.
What this actually means
No single model wins across the board, and that's the finding, not a gap in the analysis. Equisect Dynamic's strong path-shape tracking doesn't come with good overall calibration — it's good at getting the relative shape of a path right while being systematically off about its overall level, a specific and different problem from "the shape is wrong." Equisect Ledger's pattern is close to the reverse: more reliable about the eventual level and about not missing badly at any one point, less reliable about the path's timing along the way. Equisect Cohort doesn't have Dynamic's or Ledger's specific strength, but also doesn't have either one's specific weakness.
One number worth sitting with regardless of which model you look at: the best cross-correlation achieved by any model here is 0.571 — a real but modest relationship, not a strong one. Even the model that tracks direction best is still a long way from closely following the real quarter-by-quarter shape of what actually happened. That's arguably the more important takeaway than any single model's ranking: path shape is a meaningfully harder thing to get right than the profitability question the platform's other out-of-sample verdicts test, and none of the five models available today should be read as having "solved" it.
How to actually use this
These four numbers are a diagnostic lens, not a verdict — there's no pass/fail threshold attached to any of them (unlike the deflation battery's DSR/bootstrap-CI/multiple-testing gates). Read them alongside band coverage and AEPC for the specific fund and cohort you're pricing, not as a replacement for either. If you're pricing a fund where the timing of cashflows matters a lot to your own liquidity planning, Equisect Cohort's or Equisect Ledger's shape-tracking profile is more directly relevant than any single blended number; if what matters most is the overall level of value, Equisect Ledger's calibration edge is the more relevant one of the two.
Research and software, not investment advice.