Inside the architecture

Four hundred datasets, one point-in-time contract

Model sophistication cannot compensate for a contaminated sample, and breadth cannot compensate for inconsistent vintaging. The platform ingests approximately 400 distinct source datasets spanning market microstructure, corporate disclosure, insider activity, national accounts, monetary and fiscal statistics, energy balances and international trade flows. Every one of them enters under the same contract: values are bound to the date they were first knowable, and no downstream consumer can read a revision that had not yet been published.

Three classes of bias are addressed at the data layer rather than corrected downstream.

One state vector, read identically by every consumer

Each symbol is represented on each trading day by a unified state vector of roughly two thousand engineered dimensions, materialised in a single feature_samples table so every downstream consumer reads identical inputs.

Nine orthogonal dimensions, fused into a single market state

Financial time series are non-stationary. A model estimated over a pooled sample implicitly assumes a single data-generating process, which is false across regime boundaries and is a principal cause of live-versus-backtest divergence.

Nine dimensions are estimated independently, each by a specialised Gaussian hidden Markov model over its own observable set — and, critically, over its own source. The credit dimension reads spreads and funding conditions; the inflation dimension reads price statistics from FRED and Eurostat; the liquidity dimension reads system-wide capital availability. Dimensions are orthogonal because their inputs are genuinely independent, not because a rotation was applied to a correlated block.

Structural diversity, not member count

The estate is deliberately heterogeneous. Ensemble benefit derives from error decorrelation, not from member count, so families are selected for structural diversity — tree-based, sequential, relational, state-space, quantile and analytical — rather than for variations on one inductive bias.

Forecasts are produced at four horizons: 1, 5, 21 and 63 trading days. The 126- and 252-day horizons were retired. Their information coefficients did not survive the promotion gate at a standard that justified the estate cost, and a horizon the platform cannot forecast honestly is better absent than present and weakly weighted.

Counts are instances under independent lifecycle management — each separately versioned, separately validated and separately promotable. The figures are a census, not a marketing aggregate; they are published here so they can be checked.

Seven voices, aggregated online, weighted by demonstrated skill

Individual model outputs are consolidated into a single belief per symbol per horizon through Bayesian online aggregation. The distinction from static model averaging is material: BOA updates its weight vector sequentially as each outcome is observed, with regret bounds relative to the best voice in hindsight. It does not require a stationary posterior over models, which is the assumption that breaks first when a regime turns. A voice that degrades loses weight within days rather than at the next retraining cycle. A voice that recovers regains it without manual reinstatement.

The two conformalised quantile voices are structurally distinct from the other five. The kinematic and relative voices are derived feature mappers — they transform pre-computed valuation state into a return estimate through a fitted mapping. The CQR voices are direct per-horizon outputs carrying finite-sample coverage guarantees. iv_cqr_prob is the only voice whose probabilistic output is accompanied by a coverage guarantee, which is why it dominates any downstream Kelly-style sizing calculation.

Per-horizon weights track each voice's empirical information coefficient profile. A Sharpe gate rebalances weights as out-of-sample observations accumulate; voices failing the gate are progressively down-weighted rather than removed.

Distribution-free intervals with finite-sample coverage

Point estimates are insufficient for capital allocation. The platform implements split conformal prediction with adaptive conformal inference (Gibbs & Candès, 2021), which provides intervals satisfying P(Y ∈ C(X)) ≥ 1 − α without distributional assumptions on the underlying process, in finite samples rather than asymptotically.

Conformal prediction in its standard form assumes exchangeability, which financial time series violate through volatility clustering, regime transition and structural break. Adaptive conformal inference addresses this by adjusting nonconformity thresholds dynamically, with separate calibration sets maintained per regime. Interval width therefore widens in high-volatility regimes and contracts in calm ones as a property of the method, not as a post-hoc adjustment.

Posterior calibration is monitored in-pipeline and daily, under probability-integral-transform uniformity testing and Brier-score gates. Calibration drift is treated as a production incident, not as a research finding.

Regime-conditional intrinsic value, re-estimated daily

Standard discounted-cash-flow valuation is estimated infrequently against fixed assumptions, which makes it structurally unable to inform positions held over one to three months. Three modifications adapt it to that horizon.

Five intelligences, separated by question

The system is decomposed by question owned, not by pipeline stage. Each service answers exactly one question and is structurally incapable of answering another's. One rule governs every boundary.

A service may publish a quantity only if it also validates that quantity.

The rule is greppable and it settles ownership disputes mechanically. Risk Intelligence backtests its own value-at-risk under Kupiec, Christoffersen and Basel procedures, so it owns the covariance matrix, factor loadings and tail parameters. No service backtests crowding as a risk quantity, so crowding does not live in Risk Intelligence — it lives in feature engineering.

Fourteen engines, each validating what it publishes

Alpha without risk control is uncompensated variance. Every proposed allocation change is evaluated synchronously against fourteen specialised engines.

Trades that cannot cover their own friction are never issued

The gap between backtested and realised performance is predominantly an execution-cost gap, and it is where most systematic strategies fail in production. Per-symbol market impact is estimated under Almgren-Chriss with Corwin-Schultz spread estimation, and modelled round-trip cost is compared against predicted statistical edge before a trade candidate is published. Where cost exceeds edge, the candidate is suppressed — not flagged, not weighted down. Suppressed.

This is a deliberately expensive design choice. It removes real signal from the output. It removes more false signal than real.

Ten statistical tests, enforced by pipeline

Model promotion is a mechanical process, not an editorial one. A candidate must clear every gate before serving a single prediction.

Supporting this, 7,050 hyperparameter optimisation trials execute weekly under walk-forward validation, with three independent A/B engines for model comparison and automated drift detection triggering retraining. Every experiment is tracked with parameter logging, metric recording and artefact versioning for full reproducibility.

What this architecture does not solve

A system's boundaries are more informative than its capabilities. These are the constraints we consider material and do not expect to remove.

The architecture is the argument.

Nothing above is a claim about future returns. It is a description of how the system is built, how it is validated and where it stops. Thirty days of full access is enough time to check whether the description holds.