01
Proof, WithProvenance.
Evidence by module — clearly sourced, historically framed, and separated from illustrative or third-party results. No blended performance strip. Every number has a source, period, and context.
Pillars
02
Validation studies
03
Methodology
04
Governance
05
Explainability
06
Research & publications
ValidationStudies
Compass — agency-MBS instantiation (legacy positioning, historical)
79.3%
Quarterly directional accuracy
1.44
Ranking Sharpe(140-period holdout)
+5.96%
CAGR vs. MBB (monthly Top-20 optimization baseline)
07
Optimization arms benchmarked
Source:
APOS Overview v2.3, PredictNow.ai, August 2026 — pre-rename, pre-Compass-repositioning. Three separate holdout tests, not comparable to one another.
Corrective Warehousing —Client Engagements
| STRATEGY / CLIENT TYPE | SHARPE BASE → WITH CAI | MAX DRAWDOWN | VOLATILITY |
|---|---|---|---|
| Vol-based strategy · Retail prop | 0.23 → 2.16 | −82% | −64% |
| VIX futures hedge · Hedge-fund overlay | −0.30 → 0.97 | −60% | −46% |
| Fixed income ($10B AUM) · Institutional FI | −0.69 → 0.87 | −59% | −50% |
| Multi-asset systematic · Multi-strategy fund | 1.11 → 1.59 | −30% | −24% |
| Thematic ETF (24 names) · Portfolio optimization | 0.75 → 1.17 | −25% | −33% |
Source:
Corrective Warehousing Proposal v2, August 2026. Rows not comparable to one another; no aggregate should be drawn.
True North —Published Academic Evidence.
Finding
Modified book vs. the manager, per quarter
Funds improved over their lifetimes
Constrained vs. unconstrained signal, per quarter
Figure
+ $17.1M | avg fees $3.6M | manager alpha $2.8M | t = 27.88
93% | paired Sharpe +0.17
1.37% vs. 3.24%
Source:
deHaan, Lee, Liu & Noh, “The Shadow Price of ‘Public’ Information,” working paper, May 2025 — 3,337 US active equity mutual funds, 1990–2020, 2022 dollars. Published academic results, not PredictNow or client performance; a lower bound under a single-investor assumption.
How We Test. Short list, plain language
Tested on your own history before anyone is asked to trust it – walk-forward, never random splits that leak the future into the past.
Evaluated on data the model never saw, and benchmarked against named published standards.
Calibration first: a probability is only useful if it means what it says.
A challenger runs alongside. If the model does not beat the incumbent on your book, we report that.
Deterministic, point-in-time tools with no language model in the scoring path.
Golden test sets, replay and drift detection. Model-vendor upgrades are measured and reported before the desk sees them.
Built To Be Reviewed. Short list
A mandate and restricted-list gate in every thread; human approval with checkpoints; an append-only audit trail.
A model card per endpoint, version-pinned scores that can be regenerated later, drift and decay monitors, and a kill switch with a fallback.
Runs inside your perimeter: no client data crosses a boundary, by architecture rather than by contract.
Every Answer Shows Its Work. Short list
Every number is grounded in a source and cited.
Every score ships with the reasons behind it, so a model-risk reviewer can see why in one sentence.
A full trace sits behind every answer: which tool ran, with what inputs, and where the numbers came from.
Research & Publications
deHaan, Lee, Liu & Noh, “The Shadow Price of ‘Public’ Information,” working paper, May 2025.
What a Pilot Result Is — and Is Not
Any result we share is a backtest or an out-of-sample research result — not a live trading return. A pilot result is a starting line, not a finish. Re-running at realised cost, adding financing and carry economics, and extending the out-of-sample window all precede any capital commitment. Where a target is missed, we say which one and by how much.
Talk through the evidence.
See the methodology, validation context and governance behind each module.