The methodology, stated plainly
Every design decision below is downstream of one objective: maximising the probability that observed historical performance generalises to unseen market conditions.
Markets are adaptive, noisy and non-stationary. That has three consequences most systematic approaches under-price. Relationships decay: a feature that predicted returns in one monetary regime may be uninformative or inverted in another. Backtests are easy to win: with enough features, horizons and parameterisations, an apparently significant strategy can always be found. Edge is fragile to implementation: a genuine statistical edge can be entirely consumed by spread, impact and turnover.
Look-ahead contamination is structurally prevented.
What data architecture lets us ask historical questions without contaminating the experiment with information that was unavailable at the time?
Most quantitative work is conducted on datasets that quietly encode the answer. Standard vendor universes contain only companies that survived. Fundamental data is commonly aligned to the reporting period rather than the filing date, giving a model information weeks before the market received it. Macroeconomic series are supplied in revised form rather than as first published.
The research environment ingests approximately 400 source datasets spanning more than 45 years: market data, the full SEC EDGAR corpus — every filing type the SEC publishes, not a curated subset — and official macroeconomic and cross-border statistics including the Federal Reserve (FRED), Eurostat, the World Bank, the International Energy Agency and international trade series. Filings are extracted period-matched and point-in-time, so a model sees a disclosure when the market saw it, not when the quarter ended.
The historical universe tracks approximately 1,775 US-listed companies from 1980, retaining bankruptcies, mergers, acquisitions and delistings: 456 of them no longer trade. Companies that failed remain in the record with the returns they actually delivered.
From that matrix, 1,319 securities are selected on institutional investability criteria — minimum price ≥ $15, average daily volume ≥ $25M, ≥ 95% trading continuity — preserving representation across all 11 GICS sectors while ensuring the liquidity that swing-horizon execution requires.
Observations are partitioned into distinct market eras, and models train only on information available at that point in history — look-ahead contamination is structurally prevented rather than avoided by discipline.
A question asked of this dataset — would this decision have worked? — returns an answer that is not quietly pre-informed by the future.
Twenty technical indicators computed from one price history are not twenty pieces of evidence.
What representation is rich enough to describe a security's condition across independent domains, without collapsing into redundant restatements of price?
Every security, every trading day, is represented as a 1,979-dimensional state vector constructed across orthogonal domains: price kinematics, order-flow and microstructure behaviour; multi-horizon trend persistence; graph-neural-network (GNN) representations of cross-asset and corporate relationship structure; corporate and cross-asset sentiment parsed with domain-specific NLP — FinBERT and the Loughran-McDonald financial dictionaries — rather than general-purpose scoring, which systematically misreads financial language; multi-scale Daubechies-4 wavelet decomposition and geometric pattern structure; period-matched fundamentals from filings; explicit Fama-French factor-exposure decomposition; macro-financial regime conditioning; and intrinsic-value dislocation measures.
Volatility structure is modelled with higher-order estimators — realised semivariance, bipower jump share, realised skew and kurtosis — rather than a single dispersion measure. Unsupervised anomaly detection (Isolation Forest) flags statistically unusual states that no supervised label anticipates. The macro layer contributes 530+ engineered multi-horizon indicators.
Breadth across independent domains means the system is not repeatedly confirming the same observation. A conviction supported by valuation dislocation, insider accumulation, factor positioning and regime alignment is a different object from one supported by four momentum indicators.
A forecast's reliability is itself a measured quantity.
Can statistically significant forward-looking signal be extracted from a high-dimensional, non-stationary state space — and can the uncertainty around it be quantified with finite-sample guarantees?
Point forecasts are the standard output of the category, and they are close to useless for capital allocation. “Expected return of 4.2%” contains no statement about dispersion, and therefore no basis on which to size a position.
Forecasts are produced at 1-day, 5-day, 21-day and 63-day horizons, matched to the swing structure of the platform.
The predictive layer uses deliberately diverse model families rather than an ensemble of near-identical gradient-boosting variants. Diversity in inductive bias is the point of an ensemble; correlated learners provide none.
Model outputs are consolidated via Bernstein Online Aggregation — an online-learning aggregator with formal second-order regret guarantees — which re-weights each model as its realised performance accumulates, rather than fixing weights at training time. This is the mechanism behind the homepage's claim: the system re-weights what it believes, every trading day.
Prediction intervals are produced through Conformal Quantile Regression — quantile bundles combined with split conformal prediction under Adaptive Conformal Inference — yielding distribution-free intervals with finite-sample coverage guarantees that hold without assuming the return distribution takes any particular form. Every model is continuously monitored on conformal interval coverage, calibration slope and CRPS (the continuous ranked probability score), so a forecast's reliability is itself a measured quantity.
Coverage guarantees are what allow uncertainty to be an input to position sizing rather than a caveat in a footnote.
Regime is a routing and interaction term — not a flat additive predictor.
Can changes in the macro-financial environment be detected early enough, and reliably enough, to alter model behaviour and portfolio risk before the change is priced?
Most systematic strategies are regime-blind. They apply one set of learned relationships across credit expansions and credit crises alike, and discover the regime change through drawdown.
The regime engine models independent axes of market behaviour — credit conditions, monetary policy, market liquidity, inflation, real interest rates, recession risk, market sentiment, volatility structure and the fundamental cycle — each estimated separately with Gaussian Hidden Markov Models, then fused into a unified market state through an overarching Hidden Semi-Markov Model. The semi-Markov formulation models state duration explicitly, which a standard HMM cannot — and regime persistence is economically meaningful.
The resulting state vector is the primary router for the platform: it conditions quantile optimisation, scales risk premia inside the daily valuation calculation, and adjusts ensemble weighting to the prevailing environment. Regime is a routing and interaction term — not a flat additive predictor.
An unqualified value signal is a value-trap generator.
Can intrinsic-value estimation adapt to changing discount rates, macroeconomic conditions and market dynamics, rather than relying on assumptions fixed at the time of the analysis?
Conventional DCF fixes WACC, terminal growth and margins at a point in time and holds them for months — precisely the inputs that move most when the environment shifts. A subtler failure: fundamental valuation runs on a slower clock than the market. A security can be genuinely undervalued and continue falling for a year; for a swing-horizon strategy, an unqualified value signal is a value-trap generator.
Guided by the HSMM state probabilities, the valuation engine re-calibrates discount structures, margin parameters and growth risk premia when a macro or liquidity transition is detected — not at the analyst's next review.
Valuation structures are fused with asset kinematics — price velocity, acceleration, localised frequency volatility — separating undervalued and turning from undervalued and still falling.
The full valuation vector is recomputed every trading day against macroeconomic releases, TIPS-implied real yields, incoming filings and current sentiment.
The edge is evaluated net, not gross.
How should security-level forecasts, each with its own uncertainty, be transformed into a coherent portfolio under competing constraints — and does the edge survive the cost of acting on it?
The most common failure in systematic investing is treating a portfolio as a list of good ideas. Ten high-conviction positions sharing a factor exposure are one position. The second most common failure is transaction-cost blindness: a backtest on closing prices can report a strong edge for a strategy that loses money in production.
Construction is organised as template → policy → portfolio: a single strategy template, fully parameterised by policy, can power thousands of independently configured portfolios. There is no house portfolio to imitate — each portfolio is constructed from the client's own risk profile, mandate and preferences, and every parameter lives in the policy, versioned and reproducible.
Portfolios are built on the belief distributions published by the predictive layer — not point forecasts — and resolved against portfolio state: multi-model Black-Litterman optimisation that respects the confidence attached to each view; sizing variants from a Kelly-derived ceiling combined with volatility targeting to conviction- and equal-weighted schemes; horizon conflict resolution; binding bounds on gross and net exposure; tiered drawdown circuit breakers; regime-adaptive overrides; and risk-profile presets, so the same research output produces different portfolios for different mandates.
Every candidate instruction carries a calibrated round-trip cost estimate built from measured effective spreads, Amihud illiquidity and Almgren-Chriss market-impact parameters — calibrated against realised execution data, not assumed. Where estimated cost exceeds the predicted statistical edge, the instruction is suppressed. The edge is evaluated net, not gross. The ≥ $25M average-daily-volume universe floor exists specifically so that impact costs remain modelable at realistic sizes.
A portfolio decision, not a ranked list — with concentration, correlation, mandate constraints and implementation cost resolved before the recommendation arrives.
A control plane, not a report.
Should this portfolio be allowed to take this change — and at what size?
Risk is not a stage at the end of the pipeline. It is a synchronous control plane: every allocation passes through it before any instruction exists, and it is the only layer whose objective function is not return.
VaR is not subadditive — a VaR budget can be exhausted by a diversifying trade. CVaR is subadditive and positively homogeneous, which yields an exact Euler decomposition of the portfolio budget into per-position contributions: “this order spent 12 basis points of risk budget” is an arithmetic fact, not an attribution heuristic. Portfolio tail risk is estimated by filtered historical simulation — chosen for identifiability at the platform's actual sample sizes, because a measured statistic with a limit can be backtested and an unidentifiable dynamic model cannot.
Extreme Value Theory with Generalised Pareto tails, paired with Clayton and Gumbel copulas, models the phenomenon whereby diversification fails precisely when it is needed — which linear correlation cannot represent. Covariance estimation pairs Ledoit-Wolf shrinkage with volatility-state conditioning rather than assuming stationarity across the window. Concentration is measured as correlation-adjusted effective N, and liquidity horizons are re-estimated under stress. Reverse stress testing is closed-form: the smallest shock that breaches a limit is computed, not guessed.
Limits live in a versioned, dated, approved limit book with four-eyes change control; regime overlays are tighten-only by construction — no override can ever widen a limit; kill-switch governance, restricted-list pre-trade compliance and an append-only audit trail are enforced in software. Risk evaluation is fully deterministic — no randomness, no wall-clock caches — so every verdict is exactly replayable.
Constraints that bind at construction produce a different portfolio from risk reports that describe one.
No single intelligence grades its own homework.
The platform decomposes the investment problem into five specialised intelligence services, fed by market data and prediction services upstream, so each function can be independently validated before recombination into a decision system:
| Intelligence | The question it owns |
|---|---|
| Investment | What do we believe? Publishes belief distributions, with uncertainty, per security and horizon. |
| Strategy | Which investment policy is optimal? Searches the policy space and validates strategy templates. |
| Portfolio | How do we express and implement? Owns construction, position lifecycle, rebalancing and implementation cost. |
| Risk | Should we allow this? A synchronous control plane operating pre-trade, intra-horizon and post-trade. |
| Experiment | How does the system improve? Adjudicates evidence and emits recommendations — and never acts. |
A service may publish a quantity only if it also validates it.
The layer that evaluates whether an edge is real holds no position in the answer — it will never promote a model, change a weight, or size a position. Two doctrines follow: gating controls consumption, never evaluation — blocked predictions are still scored, which is precisely what makes every gate falsifiable; and limit changes are themselves gated through pre-registered hypotheses, so even the risk framework's own evolution requires declared, testable evidence.
Because each stage is independently validated, a failure can be localised to a layer — a selection problem is distinguishable from an allocation problem from an implementation problem — rather than diagnosed as “the model is underperforming.”
The platform runs a pre-registered experiment on itself, every trading day.
Model skill is evaluated on a Model × Context lattice — roughly 30,000 evidence cells per day — across sectors, liquidity buckets and horizons, with empirical-Bayes shrinkage so that a thin cell is pooled toward the prior and cannot over-claim: an apparently spectacular result on eight observations shrinks to nearly nothing, and says so.
Rolling statistics flatter themselves: a rolling IC series can carry lag-1 autocorrelation near 0.97, so months of daily values contain only a handful of independent observations. The evidence layer therefore derives non-rolling, per-day metrics and reports effective sample sizes computed from them — a deliberately harder standard.
Every monitored question is a standing, pre-registered hypothesis: metric, direction and minimum effect are declared before any evidence is read, and pre-registration is enforced in software, not advisory. Verdicts are corrected for multiplicity (Benjamini-Hochberg across each day's hypothesis family), inference uses block-bootstrap methods that respect serial dependence, and robustness is tested with the Deflated Sharpe Ratio in both directions — because a low probability of skill is evidence of skill not established, not of harm demonstrated, and the two must never be conflated.
Page-Hinkley change detectors monitor every model-horizon pair for degradation, and distribution shift is vetoed with two-sample Kolmogorov-Smirnov tests — chosen over population-stability heuristics whose thresholds were calibrated on samples three orders of magnitude larger than a trading day provides.
The system that evaluates whether an edge is real never trades on the answer — and every claim it makes was specified before the data arrived.
The evolution is real — and it is disciplined.
The homepage says the platform evolves. This section states, precisely, what that means — because “self-learning” is the most abused claim in this category, and precision is the difference between an adaptive system and an unaccountable one.
| Layer | Update mechanism | Cadence |
|---|---|---|
| Feature matrix | Full recomputation across the production universe | Every trading day |
| Valuation estimates | Re-estimation against current macro, yields, filings, sentiment | Every trading day |
| Regime state | Re-inference of the market state vector | Every trading day |
| Belief consolidation | Bernstein Online Aggregation across model outputs | Continuous, per observation |
| Model weights & architecture | Retraining and promotion under statistical gates | Periodic, gated, never automatic |
| Strategy templates | Authored, validated, then promoted or rejected | Periodic, gated |
Beliefs, regime state and valuations adapt continuously as outcomes arrive. Models and strategies change only through the explicit statistical gates of section 11. Nothing is promoted into production by the system itself. The loop is continuous in evaluation; it is deliberately discontinuous in deployment.
The difference between a research programme and a backtest.
How do we distinguish a real edge from the best of many things we tried?
If a hundred strategy variants are tested, several will show a t-statistic above 2.0 by construction. A conventional significance threshold applied to a mined search space is not evidence. Candidate models and strategies must therefore clear a battery designed specifically for mined search spaces:
| Gate | Standard |
|---|---|
| Multiple-testing correction | Verification t-statistic ≥ 3.0 under the Harvey, Liu & Zhu framework, alongside Benjamini-Hochberg / Benjamini-Yekutieli / Holm family control |
| Data-snooping tests | White's Reality Check, Hansen's SPA and Romano-Wolf stepdown across the full trial family |
| Sharpe deflation & track sufficiency | Probabilistic and Deflated Sharpe Ratio, minimum track-record and minimum backtest length — with the number of trials counted honestly, never assumed to be one |
| Overfitting probability | Combinatorially symmetric cross-validation (CSCV) estimate of the Probability of Backtest Overfitting |
| Resampling done right | Stationary block bootstrap at the Politis-White optimal block length; combinatorial purged cross-validation with embargo against leakage across folds |
| Search discipline | Hyperparameter search (Optuna TPE) evaluated on a lower-confidence-bound objective, with stability required across 7,050 weekly walk-forward trials |
Kill-switch criteria — a required out-of-sample floor — are declared before the result is seen; declaring the threshold in advance is what makes it a test rather than a rationalisation. Failing any gate returns the component to research.
One component of an evidence framework, not the argument itself.
Performance is presented here — at the end of the technical case, not the beginning — because it is one component of an evidence framework, not the argument itself.
A deliberately mechanical track that isolates raw predictive selection: the week's top-10 long and top-10 short by model belief, rebalanced after Monday's close, with no portfolio optimisation. Because nothing downstream can flatter it, it is the cleanest public measure of whether the models select well.
| Track | What it isolates | Construction | Result |
|---|---|---|---|
| Golden Tickets | Raw predictive selection | Weekly top-10 long & short, no optimisation | Reading the published baskets… |
Transaction costs are deliberately not modelled here, and that is what makes the track readable. Golden Tickets exists to answer one question — do the models select well — so anything layered on top of the selection would blend prediction quality with an execution assumption, and a reader could no longer tell which of the two moved the figure. Implementation cost is priced where it changes a decision: inside portfolio construction, against the specific size and liquidity of a real book, where the edge is evaluated net.
There is no model house portfolio to advertise. Each client portfolio is constructed from its own risk profile and preferences — and each inherits the full evidence chain: deterministic walk-forward replay of its policy, factor attribution against its benchmark, and exact CVaR budget attribution per order. The evidence framework evaluates your portfolio's construction, not a showcase's.
A months-long live window is sufficient to demonstrate that a system operates as designed under real frictions. It is not sufficient to establish a statistically significant edge, and we do not present it as if it were.
Any published performance figure ships with: the exact observation window (start and end dates); gross and net of all costs, fees, commissions and slippage; annualised volatility and maximum drawdown; Sharpe or Sortino ratio; average and maximum gross and net exposure; beta and factor attribution versus benchmark; turnover; and the benchmark stated as total or price return.
The honest formulation — and the only one we use — is: “observed during our live validation period, under the methodology and disclosures set out here.” Never “proven to outperform.”
Past performance does not guarantee future results. Nothing on this page is investment advice. All investing involves risk of loss.
Investment intelligence begins where prediction ends.
No credit card required · You keep final authority over every action.