What the workspace can compute
The method catalogue spans the whole research surface: data quality checks, statistical tests, time-series and volatility models, validation schemes, portfolio optimisers, option pricing, factor models and reporting. Only a slice of it is predictive modelling. The 43 machine learning models are the distinct, selectable estimators a researcher actually picks from on the ML Models page, and they carry their own training, validation and explainability path.
01What you actually do here
Classification or regression, then a named estimator from the registry — logistic regression through XGBoost and an MLP classifier on the classification side, linear regression through gradient boosting variants on the regression side. Each entry states whether it is seeded for exact reproducibility.
Chronological split, train-only scaler fit, no shuffling and a purged embargo are on by default. The purge gap is floored at the label horizon, not just the embargo you set, because the last rows of training share their outcome with the first rows of test regardless.
Rolling-origin cross-validation, five folds by default, refitting the scaler inside every fold so the transform cannot leak across it. The model trains fresh on this run’s data and parameters; nothing is served from a cached fit.
Model Tuning searches one estimator’s hyperparameter grid and reports whether the fit generalises. Experiment Comparison ranks competing estimators on the same data, split and threshold, sorted by Deflated Sharpe rather than raw Sharpe. Model Comparison itself exercises three fixed classifiers — logistic regression, random forest, XGBoost — not the full fourteen.
Permutation importance and partial dependence, both model-agnostic, computed out of sample on the same leakage-safe pipeline the model was trained on — not on the training data, where importance would flatter features the model merely memorised.
43 models, 5 families, one training path
Hover to pause
01bThe catalogue as a constellation
Each point is a method from the workspace's catalogue, clustered by category. Hover to read its name; choose a category to isolate it. Filled points have been checked against a published worked example, a closed-form identity or an independent library. Outlined points have not yet, and the workspace says so.
Categories
Select a category, or hover a point.
The count beside each category is methods checked against a reference over methods in the category.
Separately, the ML Models page offers 43 machine learning estimators: 14 classifiers, 16 regressors, 6 clustering methods, 4 dimensionality reducers and 3 anomaly detectors. The catalogue and the model registry are different things, and the site never adds them together.
01cEvery name, searchable
Filter the research catalogue by category and validation status, or the ML registry by family. Click a name to copy it. The filtered view is in the address bar, so it can be shared.
119 of 119 methods
Optimal trade scheduling under linear temporary and permanent impact. Trades faster when risk aversion is high, paying impact to avoid price risk; at zero risk aversion the schedule is exactly TWAP.
P&L of the spread position accounted through both legs: one unit of signal is long 1 y and short β x, returns are Δspread over the gross capital |y| + |β||x|, and costs are charged on the notional actually traded at each position change. Publishes the gross series and turnover so the cost-sensitivity sweep can re-price the same trades.
Processes each bar in order with an order queue and a cash ledger. Orders are queued on the bar that decides them and filled on the next, compete for that bar's volume, and are paid for out of cash — so an order can fill partially or not at all. Slower than the vectorised engine and the only one that models any of that.
Impact proportional to the square root of participation. The standard practitioner approximation. The order is described by the parameters below: this workspace records no orders, and the entry used to declare that it consumed one, receive none, and quietly report the library's own default.
Re-runs across cost multiples to find the level at which the edge disappears.
Signal-to-equity simulation over a return series. Fast; assumes fills at the modelled price.
Flag price jumps consistent with an unadjusted split or capital event. Detection only — a jump matching 1:2 is also what a genuine 50% fall looks like, and adjusting on that evidence would rewrite real history on a guess.
Bring prices, returns, factor returns, benchmark weights, fundamentals, execution records or macro series into a run, recording which provider supplied them and a hash of exactly what arrived.
Stamp every record with the date it was actually knowable: the later of its stated availability date and its period end plus a declared lag.
Convert a series between frequencies with an explicit aggregation rule. Refuses to resample a return series with an averaging rule: returns compound. Every later step in the workflow reads the resampled series; the original rows stay available as data_prices_original.
Refuse to proceed while blocking data problems remain. Reads the report an earlier quality.assess step produced and fails the step when anything at or above the chosen severity is still open, which blocks every step that depends on it and leaves the rest of the workflow to run.
Missing values, duplicate and out-of-order timestamps, timezone mixing, calendar gaps, stale prices, extreme moves, unadjusted actions, survivorship and look-ahead risk.
Delta, gamma, vega, theta and rho in closed form, with the units stated per Greek and the practitioner scalings (per vol point, per calendar day, per percentage point) carried beside the raw derivatives — a Greek whose units are unstated is a units error waiting to be booked. Refused at expiry or zero volatility, where delta is a step function and any finite number would misrepresent a discontinuity.
European option price in closed form, with the continuous dividend yield of Merton (1973). Degenerate cases return the exact limit — intrinsic value at expiry, the discounted certain payoff at zero volatility — never NaN.
Zero-coupon bond price under Cox-Ingersoll-Ross dynamics, face 1, in the affine closed form with log-space A. Refused at sigma = 0, where the closed form's exponent is undefined — the deterministic limit is stated in the refusal instead of silently substituted.
Cox-Ross-Rubinstein binomial pricing, European or American. American exercise compares continuation with intrinsic value at every node, and the output separates the early-exercise premium from discretisation error. Refuses a lattice too coarse to mean anything and a discretisation whose risk-neutral probability leaves (0, 1).
Estimates dr = kappa (theta - r) dt + sigma sqrt(r) dW by OLS on the variance-stabilised Euler discretisation. An approximation, consistent as dt -> 0, and the result says so; refused outright for a series touching zero, where the square-root diffusion is undefined.
The volatility that makes Black-Scholes-Merton match a quoted price, by Brent root-find on a function that is strictly increasing in volatility. The quote is typed in: this workspace ingests no option chains, so there is no market feed to read it from, and a whole smile or surface is out of reach until one exists. A price outside the no-arbitrage bounds is refused with the violated bound named, never mapped to a number.
American options by least-squares Monte Carlo: seeded risk-neutral paths, continuation value regressed on (1, S/K, (S/K)^2) over in-the-money paths, exercise where intrinsic beats the fit. The in-sample estimator is biased low from policy suboptimality and the output says so beside its Monte Carlo standard error.
European option under geometric Brownian motion plus lognormal Poisson jumps — the Poisson-weighted sum of Black-Scholes-Merton prices. Relaxes exactly the pure-lognormal assumption the volatility smile contradicts. With no jumps it reduces to Black-Scholes-Merton exactly, and the suite asserts it.
Estimates dr = kappa (theta - r) dt + sigma dW from an observed short-rate series by OLS on the exact AR(1) discretisation, which is the exact conditional maximum-likelihood estimator. A sample showing no mean reversion is refused rather than reported with an imaginary half-life.
Zero-coupon bond price P = A(T) exp(-B(T) r0) under Vasicek dynamics, face 1, with the continuously compounded yield beside it. Prices off stated parameters or off a fit from earlier in the run — not off a market curve, because the workspace ingests none.
Total cost against the decision price, decomposed into delay, impact and opportunity cost on the unfilled remainder.
Splits an order evenly across time. Ignores the volume profile entirely, so it over-participates when the market is quiet — which is exactly when impact is highest.
Splits an order in proportion to expected volume, which holds participation — quantity over volume — constant in every period. Refuses without a volume profile: VWAP with a flat profile assumed is TWAP under a misleading name.
Splits active return into allocation, selection and interaction. The two variants differ per group and agree in total when both weight sets sum to one.
Fama-French three plus momentum.
Build SMB, HML, RMW, CMA and the market factor by independent double sorts on the project's own universe: size at the median, the characteristic at the 30th/70th percentiles, value-weighted portfolios, sort variables strictly as-of the formation date. Every deviation the data forces — no NYSE breakpoints, no size dimension without market capitalisation — is recorded on the result, and missing fundamentals are refused rather than proxied silently.
Frazzini-Pedersen BAB on the project's own universe: trailing betas shrunk toward one, rank weights w = k(z − z̄)^±, the low-beta side levered and the high-beta side delevered to a beta of one, so the factor is beta-neutral ex ante by construction.
Winners minus losers on each security's own prior return from t−window to t−skip — 12-1 momentum at the daily defaults. With market capitalisation and value weighting this is the published 2x3 size/prior-return UMD; without, it is a univariate equal-weighted sort and the result says so.
Explains returns by market, size and value, optionally with profitability and investment. Reports Newey-West standard errors alongside the ordinary ones, because autocorrelated residuals make an ordinary alpha t-statistic too large.
Factor **return series** extracted from the universe's return panel rather than specified in advance, with the loadings that produced them. This entry used to point at a route that runs PCA over four summary statistics per ticker — annual return, volatility, Sharpe, drawdown — which has no time dimension and so cannot produce a factor return series at all.
Where a security stands against its peers on the same date: rank, percentile, z-score within the cross-section, momentum relative to the cross-section and volatility against its median. Nothing else in this workspace computes a value that depends on how a security compares with anything but its own history, so a model could learn that a stock rose 3% and never that it was the worst in the universe that day.
Difference a series enough to be stationary while keeping memory.
Valuation, profitability and leverage ratios from reported statements, built only from a point-in-time aligned panel.
Given a side a primary model has already decided, was the bet worth taking net of cost? Meta-labelling does not choose direction; it learns when to act on a direction already chosen, which is a different and usually easier problem than predicting the sign, and it turns a signal's precision into something that can be raised without changing the signal.
Amihud illiquidity, Roll's effective spread and Kyle's lambda, estimated from daily prices and volume. These recover the price of liquidity without a quote book, which is what a research workspace usually lacks.
Simple and log returns, lagged and compounded returns, momentum with a skip, rate of change, price against its moving average, and the Donchian channel with its breakout flags. Every value uses only information available at or before its own timestamp.
Band positions on a strictly trailing z-score of the spread, with hysteresis: enter beyond the entry band, exit once z falls back inside the exit band. The z-score never uses the full-sample mean — that would let every early signal see the end of the sample.
Rolling mean, standard deviation, z-score, bias-corrected skewness and excess kurtosis, autocorrelation at a lag, the rescaled-range Hurst exponent and the Shannon entropy of the window's distribution.
Moving averages, Wilder's RSI, MACD, Bollinger bands, the stochastic oscillator and the average directional index — each computed from trailing data only, and each returning every component rather than one headline line.
Which of three things happened first: a profit-taking barrier, a stop-loss barrier or the holding period running out. Both price barriers are sized off the security's own trailing volatility, so the same label means the same thing in a calm period and a violent one. A fixed-horizon label says what the return was after ten days whatever happened in between; a position does not experience that.
Rolling and exponentially weighted volatility, average true range, realised volatility, downside volatility about a threshold, and the ratio of short-window to long-window volatility.
Sequentially fitted trees. Usually the strongest tabular baseline.
The baseline every other model has to beat. Coefficients are interpretable.
Multi-layer perceptron. Needs more data than most financial samples contain.
Bagged decision trees. Robust to scaling and mixed feature types.
Margin-based classifier or regressor. Sensitive to scaling.
Autoregressive integrated moving average, optionally seasonal.
Exponentially weighted volatility. Simple, fast, and hard to beat short-horizon.
Error-trend-seasonal smoothing, including Holt-Winters.
Conditional volatility, including asymmetric variants that treat falls and rises differently. The order is fixed at (1,1) and the innovations are Gaussian: this estimator implements one lag of each under a Gaussian quasi-likelihood, and the p, q and error-distribution controls that used to appear here fitted a GARCH(1,1) under whatever name was chosen.
A local-level or local-linear-trend state space, filtered forward and smoothed backward, with the state variances estimated by maximum likelihood. Both paths are returned and only the filtered one is tradeable: the smoothed level at any date conditions on the whole sample.
Fit ds = θ(μ − s)dt + σ dW to the spread by its exact AR(1) discretisation: θ = −ln(φ)/Δt, half-life = ln 2/θ, closed form. The half-life is the natural entry/exit calibration for a mean-reverting spread — a holding period much shorter than it pays costs the reversion has not yet earned back.
Joint dynamics of several series, with orthogonalised impulse responses and information-criterion lag selection. The forecast now covers every series; the first series alone remains at detail.forecast_first_series.
Density-based clustering of securities that leaves outliers unassigned.
Soft clustering with per-component covariance. Returns the responsibilities — the probability each security belongs to each component — and not only the most likely label: a security at 55% is not in a component the way one at 99% is, and discarding that is the difference between a mixture model and k-means with more arithmetic.
Nested clusters of securities from a linkage tree, cut by distance rather than by a count. No k chosen in advance — and where a k was hardcoded at half the universe and shown to nobody, the tree is now cut at the largest gap between merge heights and the number of clusters that produces is reported.
Partitions the universe's securities into k groups by distance between their return histories. Each row of the input is one security, not one observation: the question is which securities behave alike.
Orthogonal components of the universe's return panel, ordered by variance explained.
The Sharpe ratio adjusted for how many configurations were tried, and for non-normal returns.
How long a track record must be before a Sharpe ratio is statistically distinguishable from the benchmark.
Adjusts a t-statistic or p-value for the number of tests performed.
The probability that the true Sharpe ratio exceeds a benchmark, given the sample's skewness, kurtosis and length. It corrects for the shape of the return distribution and the length of the record — not for how many configurations were tried, which is the Deflated Sharpe Ratio.
The chance the selected configuration underperforms the median out of sample, estimated by combinatorially symmetric cross-validation.
A more powerful successor to the Reality Check. Studentised, and recentred so that adding hopeless strategies to the set does not make the test easier to pass.
Tests whether the best of many strategies beats a benchmark once data snooping is accounted for. Takes the whole set that was tried, not the winner.
Annualised return over the absolute maximum drawdown. Null when the series never drew down.
Annualised active return over annualised tracking error, measured on the observations the two series share.
Annualised return above what beta exposure to the benchmark explains.
Probability-weighted gains over losses about a threshold. Above one the gains outweigh the losses. Null when no period fell below the threshold.
Annualised excess return per unit of annualised total volatility. Null, not zero, on a series with no volatility.
Excess return over downside deviation, where downside deviation is the root mean square of the shortfall below the minimum acceptable return taken over **all** periods — not the standard deviation of the losing subset about its own mean, and not an average over the losing periods only. Null when nothing falls below the MAR: the denominator does not exist then, and zero is the worst value the ratio can take rather than an absent one.
Annualised excess return per unit of systematic risk — excess return over beta. Rewards a concentrated book that a Sharpe ratio would penalise, because it counts only the risk the benchmark explains.
Participation in rising and falling benchmark periods, separately. Down capture above 100% is the case worth looking at.
Blends equilibrium returns reverse-optimised from a starting portfolio with explicit views and their confidence. Those equilibrium returns are market-implied only when the starting portfolio is the market: supply capitalisations and it is, and without them equal weights stand in and the output says so. Reverse-optimising equal weights gives the equilibrium of a universe in which every security is the same size, which is not a market anybody trades and must not be described as one.
Minimises expected shortfall rather than variance. Linear-programmable.
Herfindahl index and the effective number of positions — how many equally-weighted holdings would be this concentrated.
How many independent bets a set of holdings actually represents, from the participation ratio of the correlation eigenvalues. Uncorrelated holdings give exactly N and perfectly correlated holdings give exactly 1. This is the number the Fundamental Law needs; substituting the holding count is how twenty correlated names come to promise 4.5x the information ratio of one.
The baseline that optimisers frequently fail to beat out of sample.
Allocates down a clustering tree, avoiding covariance inversion entirely.
Growth-optimal position size. Full Kelly is too aggressive for most mandates.
The efficient frontier. Highly sensitive to estimated expected returns.
Runs a rebalancing schedule across a universe and returns the portfolio's return series. Weights drift with relative performance between rebalances, and turnover is measured from the drifted weights to the new target — measuring it from the previous target charges a book that was left alone for a full round-trip every period.
Weights so that each asset contributes equally to portfolio risk. The solution is long-only and fully invested by construction, so the minimum and maximum weight controls that used to appear here had nothing to constrain and were never passed to the solver.
Counts the trials that choosing the holdings represented. A single-asset validation counts parameter settings and cannot see the securities that were examined and dropped, so the deflated Sharpe is computed against a trial count that omits the largest search actually performed.
Latent states with their own means and variances and transition probabilities between them, fitted by Expectation-Maximisation. Returns the per-state parameters, the transition matrix, the filtered and smoothed state probabilities and the most likely path. What this entry pointed at until now was a rolling standard deviation split at its own sample median: no likelihood, no transition matrix, no state probability, and boundaries taken from the whole sample so every label looked ahead.
Regime labels from trailing realised volatility. Each label uses only the trailing window, so the series can condition other work without leaking.
The full report: verdict, methodology, statistical evidence, data and features, strategy, assumptions, out-of-sample results, critique and next steps, reproducibility and limitations. Verdict first, so a reader who stops after the first section already has the conclusion and the reason for it. A workflow's report step used to emit three loose blocks under a summary claiming seven sections, while the real generator was reachable only over HTTP.
How many configurations were evaluated, by which search, and whether their return series were kept — which is what deflates everything else in the report. Reads the run's ledger; it shared a handler with the evidence report and never read one.
Sensitivity to a benchmark and the volatility of the difference, measured over the overlapping observations the two series share.
The average loss in the tail beyond VaR. Coherent, unlike VaR.
Maximum, average and duration of peak-to-trough declines.
Fits the tail directly, by a generalised Pareto distribution over the peaks beyond a threshold, rather than assuming a distribution for the whole return series.
The loss threshold exceeded with a given probability over a horizon.
Tests for conditional heteroskedasticity — volatility clustering.
Tests for a unit root. The null is non-stationarity.
Finds an unknown number of breaks at unknown dates by dynamic programming, which is globally optimal for each number of breaks — the best two breaks are generally not the best one plus the best of the remainder, so a greedy search gets this wrong.
Detects parameter instability from cumulative residuals.
Tests for a coefficient break at a known date.
Build the actual spread series y − a − βx from the hedge ratio an upstream cointegration step estimated. The ratio is consumed, never re-estimated: the spread must be the same residual whose stationarity the test certified.
Two-step test for a cointegrating relationship between two series. Publishes the first-stage hedge ratio and the exact level series it was estimated on, so spread.construct can build the spread the test was actually about rather than re-deriving a slightly different one.
Tests whether one series improves the forecast of another. Predictive precedence, not causation.
Estimates long memory. Above 0.5 suggests trending, below 0.5 mean reversion.
Tests normality through skewness and kurtosis. Financial returns usually fail it.
Tests for cointegrating rank among several series at once, by the sequential trace test. Reports the whole ladder, not just where it stopped, so a marginal stopping decision looks marginal.
Tests for stationarity with the opposite null to ADF. Run both: they disagree informatively.
Tests whether autocorrelations up to a lag are jointly zero.
Tests the random-walk hypothesis across horizons.
Replays the returns the securities in the book **actually printed** between the dates of a named episode. The dates are facts and are shipped with the workspace; the returns are read out of the loaded price history and are never modelled. When that history does not span the window the method refuses and says what it covers — because a scenario named after a real period whose returns were invented is worse than no scenario. Until now the six 'episodes' were hard-coded shock, volatility and correlation triples in the front end, fed to the hypothetical path.
A market move you specify, propagated through each asset's beta to the benchmark, with the covariance rescaled by a volatility multiple and correlations pushed toward one. Runs through the same risk engine as the Scenario page rather than a second calculation that could disagree with it. The inputs are yours: this is not, and does not claim to be, what any period did.
Starts from an unacceptable loss and finds what would cause it.
Many train/test arrangements rather than one, each purged and embargoed. Produces a distribution of out-of-sample paths.
How many independent observations a set of overlapping labels is actually worth, and the sample weights that correct for it. Purging stops the training set overlapping the test set; it does nothing about the training set overlapping itself. Two hundred and fifty ten-day labels in a year are built from 250 overlapping windows and are worth roughly 25 independent observations, while every standard error computed from them assumes 250.
Simulates many forward return paths under a stated process and reports the distribution of terminal wealth, worst drawdown and realised Sharpe ratio. This is path simulation. It is not sampling portfolio weights — the endpoint this used to name is a portfolio-weight sampler, which answers a question about allocation — and it is not reshuffling the order of historical trades, which destroys the serial structure of the series.
Every robustness statistic that can be computed from what this run holds, assembled with the verdict that follows from the set. The verdict is a property of the whole panel — 'nothing failed but two checks are marginal' cannot be read off any single statistic — which is why it is assembled here rather than by whichever overfitting step happened to run first.
One train and one test period. One out-of-sample number, which is one draw from a distribution — and it is that number this runs. It used to run combinatorial purged cross-validation instead and ignore its own training fraction, because the split mode came from a parameter none of these three declared.
Resamples blocks so serial dependence survives resampling, and reports confidence intervals for the Sharpe ratio and the maximum drawdown. The valid alternative to reshuffling trades, which destroys the serial structure and answers a different question.
A sequence of out-of-sample periods, refitting at each origin. Rolling keeps the training length fixed; expanding grows it. It produces splits and not paths: a fold covers part of the sample, a path covers all of it, and only a combinatorial design has enough arrangements to stitch one.
02How it works
Validation status and model configuration are both derived, not asserted. Neither a method’s "validated" badge nor a model’s hyperparameters are something a page just states — both trace to a rule in the code.
03What you get out
Every model run produces the same shape of record, whatever estimator was chosen, so a classifier and a regressor can be compared on the same terms as the pipeline that trained them.
04What this does not do
No SHAP, no LIME
Explainability is permutation importance and partial dependence, both model-agnostic. Neither SHAP nor LIME is implemented in the codebase.
No deep learning beyond an MLP
The neural-network entries in the registry are multi-layer perceptrons — an MLP classifier and an MLP regressor. There is no convolutional, recurrent or transformer architecture in the registry.
47 methods are still unverified
The catalogue shows every method’s status honestly. 47 of the 119 have not yet been checked against a reference case and are labelled unverified rather than silently presented as equivalent to the 72 that have.
Models train fresh every run
Nothing is served from a stale cached fit. A model trains on the data and parameters of the run that requested it, every time.
Keep going
Related capabilities
Beta researchers pick a task, choose an estimator, and read the fold-by-fold record it actually produced.