Research

Dated, citable technical research from the RiskState engine: papers, validation reports, and technical notes for the quant and academic audience, written with the rigour those readers expect.

Cadence

• One paper every six months — full-length validation report covering a scoring version, an out-of-sample window, or a specific structural question. Written in academic format: abstract, methodology, data, results, limitations, citation.
• One technical post per month — shorter, focused notes on single findings or methodological decisions.

Relationship to the rest of the docs

Findings published here reference the Backtest framework and Methodology page for context; they do not duplicate them.

Editorial discipline

• Findings are reported with confidence intervals, effective sample size, block sensitivity and exclusion counts. Single point estimates without error bounds are not published.
• Pre-specified contrasts are stated before running the test. Post-hoc quintile mining is noted explicitly when it occurs.
• Scoring-version and out-of-sample gates are respected: findings from an in-sample window are labelled as such and not extrapolated.
• Methodology open, calibration closed. Enough detail is published that a reviewer can replicate the analysis against their own snapshot dataset; scoring weights, thresholds and coefficients are not.
• Every paper ships with a BibTeX-ready citation block at the end.

Publications

Papers

• Backtesting the Policy, Not the Signal — 11 June 2026. First exposure-matched audit of the engine's sizing output (max_size_fraction) rather than its score. Three benchmarks held to the policy's own mean exposure (static, vol-target, buy-and-hold context) on 275 live policy decisions (Feb–Jun 2026) plus an 8-year public-history reconstruction (2018–2026, BULL/BEAR/RANGE). Result: the policy's level of exposure dominates loss containment vs buy-and-hold (7d deep-tail frequency 3.4% vs 23.8%); its dynamic component does not yet beat a constant cap at 7 days and loses on net to exposure-matched volatility targeting, while clearly helping at 24 hours. Diagnosis (coefficient elasticity + signal-redundancy audit) and the resulting score_v4 candidate — realized-volatility-first sizing — are pre-registered ahead of the 1 October 2026 decision gate, with live shadow telemetry running since 11 June. score_v3.2 freeze through 19 November 2026 maintained.

• First Out-of-Sample Check of the Composite Quintile Contrast — 18 May 2026. First validation check. Pre-specified Q2∪Q3 vs Q4∪Q5 win-rate contrast on 7-day BTC + ETH returns, circular block bootstrap, IS/OOS split at the 8 April 2026 cutoff. 90-day snapshot dataset, 35-day matured OOS window. Pre-registered internally on 2026-04-08 (audit on request). Primary score_v3 rescored result: pooled OOS gap −3.2 pp, 95% CI [−30.7, 23.9]; per-asset decomposition reveals BTC/ETH direction disagreement out-of-sample (BTC −6.7 pp, ETH +1.8 pp). Under the framework's pooled-only-when-agreed policy the pooled OOS headline is suppressed. Classification: OOS weakened, not disproven. score_v3 freeze through 22 October 2026 maintained.

Subscribe or follow @riskstate_ai to be notified when subsequent posts land.