Private beta. A research and analytics workspace — not a trading or execution platform.

Quantitative research,
composed.

One workspace where you bring your own data, compose research methods into a pipeline, validate with purged walk-forward folds by default, and export a result you can defend. Configuration, not setup code.

Get early accessSee how it works
Research methods in the catalogue
119
Checked against a published reference
72 of 119
Machine learning models, one training path
43
scroll

01The problem

Six tools. No memory.

Most quantitative research still runs like this: a data export, a spreadsheet to clean it, a notebook to model it, a stats library to test it, a charting tool to look at it, and a document to write it up. Each handoff is a manual copy. The universe used in one analysis quietly differs from the next; the backtest runs on data cleaned differently from the features; a Sharpe ratio is quoted without the number of trials it took to find it.

The methods are not wrong. Nothing connects them, so nothing is reproducible, and a week of plumbing sits between a question and its first honest answer.

  • 1Time lost to setup that has to be rebuilt for every question
  • 2Results that cannot be re-run, so cannot be checked
  • 3Numbers that arrive in a report with no record of how they were made
Project: isolated universe, benchmark, date window, lineage, runs, reportsDataMethodsPipelineAnalysisReportSaved resultsone record of what ran, re-run on new data, every number traceable

Drag the handle. The same research, first as six tools with manual handoffs, then as one connected sequence inside a project.

02The research loop

Eight stages. One project carries all of them.

These are the workspace's own stage names and its own reasons for each. Scroll: the screen on the right is the real page for the stage on the left, panned as you scroll so the whole page can be read.

  1. Stage 01 of 08

    Universe

    Pick the index, provider and date range, then apply the securities you want.

    Every downstream statistic inherits this choice, including its survivorship bias.

  2. Stage 02 of 08

    Explore

    Return, volatility, drawdown, distributions and correlations for every security.

    What a model will be asked to explain is worth looking at before anything is fitted.

  3. Stage 03 of 08

    Data

    Align the trading calendar, handle gaps, and set the chronological split.

    A shuffled split leaks future rows into training and inflates every metric that follows.

  4. Stage 04 of 08

    Portfolio

    Weights under constraints, then shocks and historical stress applied against them.

    A strategy that survives its backtest still has to survive its portfolio.

  5. Stage 05 of 08

    Features

    Build, preview and select the predictive features.

    Feature choices are hypotheses about market structure; the model only tests them.

  6. Stage 06 of 08

    Models

    Train, forecast, tune and compare, then see which features drove the predictions.

    A model is a candidate, not an answer: compare it against others and against doing nothing.

  7. Stage 07 of 08

    Strategy and backtest

    Turn signals into rules, backtest net of costs and slippage, then watch the strategy over time.

    Classification accuracy and financial value are different questions. Ask both.

  8. Stage 08 of 08

    Evidence

    Assemble the evidence into a document, consult the method library, and keep every action and run traceable.

    A number you cannot trace back is a number you cannot defend later.

Market Explorer

Statistics & EDA

Data Preparation

Portfolio Optimisation

Feature Engineering

Tune & Experiment

Backtesting

Reports

03Inside the workspace

Real screens. Real results. No mock-ups.

Every image on this site is a capture of the workspace, never a mock-up: most from a FTSE 100 project on Yahoo Finance daily data, the rest from the bundled sample project on synthetic reference data. Click any of them to see the full screen.

One connected view of research readiness

A dashboard that reflects what is actually ready to run.

Workflow maturity, the active universe, selected features and the market snapshot are read from the project's real state, not a summary written once. The suggested next action comes from what has been run this session, and the evidence list names the stages a report can stand behind.

  • Workflow health across all eight stages
  • Suggested next action generated from this session's runs
  • Evidence in this session: which claims trace to a run
Dashboard The project dashboard once every stage has run: workflow maturity, the active universe, period movers and the market snapshot.

A real registry, not a marketing number

Every model in the registry is implemented, and every one trains the same way.

Thirty supervised estimators and thirteen unsupervised methods share one chronological pipeline: split in date order, scaler fitted on the training block only, no shuffling, a purge gap floored at the label horizon. Switch any safeguard off and the page says, in plain language, what leaks.

  • Chronological split, train-only scaler, purged embargo: on by default
  • Walk-forward folds reporting mean, standard deviation and the worst fold
  • Each safeguard explained inline, with a warning when relaxed
ML Models Walk-forward validation: five folds with a five-row embargo, mean ROC AUC 0.5546, spread 0.0472, worst fold 0.5034, and the interpretation.

Honest results, even when a model does not help

Forecast accuracy is scored against a naive baseline, whichever way it goes.

Model Forecasting predicts each held-out day one step ahead from real observations and compares the error with the simplest forecast there is: tomorrow equals today. The skill score is shown whether it is positive or not. In the captured run the model does not beat the baseline: a skill of −3.1%, printed in red and reported as exactly that.

  • Out-of-sample RMSE, MAPE and directional accuracy on a held-out window
  • Skill versus naive: 1 − RMSE ÷ baseline RMSE
  • Actual against predicted, charted so a weak result stays visible
Model Forecasting Model Forecasting for AAF.L with a ten-lag linear model: out-of-sample RMSE against the naive baseline, a skill of −3.1%, directional accuracy, the forward projection and held-out actual against predicted.

04The method catalogue

119 research methods. Every one named, 72 checked against a reference.

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.

Browse the full catalogue

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.

05Compose your own pipeline

Not a drawer of calculators. A sequence you build once and run again.

Choose methods from the catalogue, order them, and watch the same validation the workspace runs before it will save a workflow. The methods, their inputs and their outputs below are the real ones; only the execution is missing.

Start from a template

Data

Features

Validation

Models

Overfitting

Portfolio

Backtest and risk

Evidence

01 · Data acquisitionLoad a dataset from a pro…02 · Integrity checksDataset integrity assessm…03 · Feature constructionPrice and momentum featur…04 · Feature constructionVolatility features05 · Sample splittingCombinatorial purged cros…06 · Supervised learningGradient boosting07 · SimulationVectorised backtest08 · Multiple-testingDeflated Sharpe Ratio09 · Evidence assemblyEvidence report
  1. 01Load a dataset from a provider
  2. 02Dataset integrity assessment
  3. 03Price and momentum features
  4. 04Volatility features
  5. 05Combinatorial purged cross-validation
  6. 06Gradient boosting
  7. 07Vectorised backtest
  8. 08Deflated Sharpe Ratio
  9. 09Evidence report

Valid

  • Every step's inputs are produced earlier or supplied by the project.
workflow: preview
steps: 9
validation: valid

01  Load a dataset from a provider  [data.fetch]
02  Dataset integrity assessment  [quality.assess]
03  Price and momentum features  [features.price]
04  Volatility features  [features.volatility]
05  Combinatorial purged cross-validation  [validation.cpcv]
06  Gradient boosting  [model.gradient_boosting]
07  Vectorised backtest  [backtest.vectorised]
08  Deflated Sharpe Ratio  [overfit.deflated_sharpe]
09  Evidence report  [report.evidence]

Preview only. In the workspace the same validation runs before a workflow is saved, and an invalid workflow is refused rather than stored.

A preview of the Workflow Composer. In the workspace, a saved workflow is versioned on every save, 5 templates ship as starting points, and every run records which steps ran, were skipped, were blocked or failed, with the parameters actually used.

06Validation and evidence

Validation that survives a sceptic, switched on before you arrive.

Written for the reader who checks the method before the result. Every rule below is the workspace's default, and every default can be inspected on the page that uses it.

5
1

purge gap = max(embargo, horizon) = 5 rows

Move the controls. The purge gap follows the workspace's rule, not the slider alone: it is never smaller than the label horizon. Four split modes are available in the research layer: single, rolling, expanding and combinatorial purged cross-validation.

  1. 01

    Chronological split

    Rows are cut in date order. A random split scores the model on rows that come before rows it trained on.

  2. 02

    Scaler fitted on the training block only

    Standardisation never sees the test rows, and is refitted inside every fold.

  3. 03

    No shuffling

    Training rows keep their order, so sequence-sensitive estimators see the sequence.

  4. 04

    Purge gap floored at the label horizon

    The label on row t uses prices up to t + h, so the last h training rows are removed whatever the embargo says.

  5. 05

    Walk-forward folds, worst fold reported

    Mean, standard deviation, minimum and maximum for every metric. The worst fold is a first-class figure in the exported report.

  6. 06

    Trials counted against you

    An append-only ledger records every configuration tried. The Deflated Sharpe Ratio refuses to compute without it.

Reproducibility checklist, as recorded on every experiment

  • Dataset identified by id and content hash
  • Every parameter recorded
  • Strategy and engine versions recorded
  • Random seed recorded
  • Library versions recorded
  • Split design, embargo and assumptions recorded

The experiment record hashes its inputs with SHA-256, so the same configuration always produces the same hash and a re-run that disagrees with its twin is caught. Validation verdicts are write-once. The trial ledger is append-only. None of this is described as tamper-proof, because it is not; it is described as what it is.

Audit & Governance Audit & Governance: the project experiment lineage, one row per recorded action with its actor, artifact and status, filterable and exportable as CSV.

07Built for how you work

A professor and a prop researcher should not get the same pitch.

Universities and students

Teach the method, not the environment. Students spend the term on research questions, and every submission can be re-run.

Is this rigorous enough for academic work?

Purged, embargoed, walk-forward validation is the default, and 72 of the 119 catalogued methods are checked against a published reference. The method library cites its sources.

  1. 01Install the sample project or upload a CSV
  2. 02Prepare, model and validate through the leakage-safe pipeline
  3. 03Export the evidence pack with the worst fold on page two

08Results you can hand over

One record of what ran. Four ways to hand it over.

The evidence pack is generated from the same description of the session in every format, and a section appears only when the work behind it was actually run. Page two carries the fold count, the mean, the standard deviation and the worst fold. Page three carries the limitations, including the line that the workspace performs no live order routing.

  • PDFFixed evidence pack, eight sections, drawn on an A4 canvas
  • DOCXThe same model as an editable Word document
  • HTMLSelf-contained, renders offline
  • JSONThe same model as data, for a downstream service

Beyond the fixed pack, the Report Workbench assembles typed blocks — tables, metrics, formulas, prose — into a saved document, and marks a block stale when its source has been re-run with different results. Saved Results keep any panel with the parameters that produced it.

Reports and saved resultsOpen the captured evidence pack (PDF)
Page 1 of 3

Page 1: universe and data, absolute risk and return, benchmark-relative risk, portfolio construction.

09What comes next

Two things on the horizon, labelled as the horizon.

Neither of these ships today. They are shown because they shape what is being built now: the method catalogue, the run record and the numeric guard already exist, and both of these sit on top of them.

AgentQ

Future

A research agent you talk to in plain language. It would compose a pipeline from the catalogue, run it under the same validation rules as everything else, and write up the result — with every number it quotes checked against the evidence it was given. The trust layer that makes that possible is already in the workspace: a model may never produce a number that appears in a report.

Quant Code

Future

A space for your own scripts inside the workspace. A custom calculation or visualisation would get the same treatment as a catalogued method: typed parameters, a page built from its declared inputs and outputs, and results that land in Saved Results and the report with their provenance intact. The strategy specification already has a Python escape hatch in its design; this is the same idea for everything else.

The roadmap separates what exists from what is planned and from what is possible, and only the first list is a claim about the software.

Read the roadmap

10Private beta

Access opens in stages. Applications are read by a person.

What beta includes

  • The full workspace: every module, method and export
  • Free during private beta
  • A direct line to the people building it

What is asked of you

  • Use it on a real research question, not a demo
  • Report refusals, weak results and anything that felt wrong
  • Keep a short research log we can read together

How access opens

  • Cohorts, not a queue: each cohort is small enough to talk to
  • An invitation arrives by email; there is no self-serve sign-up yet
  • Universities and teams are onboarded through a separate enquiry

Start an application

Two fields here, six more on the next page. About three minutes.

Nothing is stored until you submit on the next page. Read the privacy policy.

11Common questions

Six questions, answered plainly.

The full list has 23 more, grouped by product, method, data and security, access, academic use and technical detail.

All questions
Is Quant Compose a trading platform?

No. It is a research and analytics workspace for the stage before an order exists. Nothing in it places trades, routes orders or connects to a broker, and its health endpoint reports live_order_routing: false. Strategy Monitor tracks strategies you have built and backtested; it is a record, not a connection to a market.

Do I need to write code?

No. Every stage is configured rather than scripted: you choose methods, set parameters and run. The method library explains what each method assumes and refuses. A custom-script space, Quant Code, is on the roadmap and labelled as future wherever it appears.

What does "119 methods" actually count?

The research layer's method catalogue: 119 named methods across 19 categories, from data quality checks and statistical tests to portfolio construction, option pricing, stress testing and reporting. 72 of them have been checked against a published worked example, a closed-form identity or an independent library. It is not a count of machine learning models; the ML Models page offers 43 estimators, and the site never adds the two together.

How does validation work by default?

Chronological split, scaler fitted on the training block only, no shuffling, and a purge gap floored at the label horizon. Walk-forward validation reports mean, standard deviation, minimum and maximum for every metric, so the worst fold is always visible. Switch any safeguard off and the page states what leaks.

Can I use my own data?

Yes, that is the intended workflow. Upload a CSV of daily prices with date, ticker, open, high, low, close and volume columns, or connect a market-data provider such as Yahoo Finance, FRED, Ken French, Alpha Vantage or Polygon, or a custom REST endpoint. Fifteen bundled index constituent lists give you a universe to start from, and every provider states its status in Settings.

How does beta access work?

Access opens in stages, by cohort, and each cohort is small enough to talk to. Apply on the beta page; an invitation arrives by email when your cohort opens. There is no self-serve sign-up during private beta.