Repetitive setup
Rebuilding the same split logic, scaler discipline and cost model for every new idea burns a day before the idea is even tested.
For quant researchers
Every experiment, on the same footing.
Same split design, same scaler rule, same decision threshold, every time. The workspace is where the record lives — comparison and validation are built in rather than reconstructed by hand before each write-up.
01A day with Quant Compose
A single research day, from a workflow template to a comparison table that names which configuration actually held up.
Hour 1
Compose from a template
Start from Signal research, Volatility modelling, Portfolio construction or Pairs and relative value, or a blank workflow. Each step names a catalogue method and binds its parameters; the composer checks every declared input before you run anything.
Workflow Composer Named research templates to start a pipeline from. Hour 2
Run across configurations
Vary a hyperparameter grid or a feature set and run it again. The trial ledger is genuinely append-only — there is no update or delete for a recorded trial, so a configuration you quietly dropped still counts against you later.
Tune & Experiment Model Tuning: a chronological grid search over C for logistic regression, the winning configuration, a generalisation verdict of mild overfitting, the C-versus-ROC-AUC curve, the leaderboard and score by fold. Hour 3
Compare on identical data
Competing estimators are ranked on identical data, identical split and identical decision threshold, sorted by Deflated Sharpe rather than raw Sharpe — the sort order that rewards the most over-searched configuration is the one this table refuses to use.
Tune & Experiment ROC curves for every candidate in the comparison: logistic regression 0.529, XGBoost 0.482, random forest 0.456. Hour 4
Read what actually drove the result
Permutation importance and partial dependence, computed out of sample on the same leakage-safe pipeline used to train — model-agnostic, not a black-box summary.
Explainability Permutation importance for the trained classifier: macd, daily return and volume change help; roc_10 and log return hurt out of sample. Hour 5
Export what ran
The evidence pack carries fold count, mean, standard deviation and worst fold for the validation, plus the parameters as actually used after defaults were resolved — the answer to what this ran with lives in the record, not in memory.
Evidence pack (PDF) Page 2: model, out-of-sample validation with fold count, mean, standard deviation and worst fold, and the strategy backtest against buy-and-hold.
02Three problems it removes
Named specifically, not as a category.
Methodology drift between experiments
A slightly different split or threshold between two runs makes their comparison meaningless, and that drift is easy to introduce by hand and hard to spot afterwards.
Incomparable results
Without a shared ledger, "I tried a few things and this one worked" is unfalsifiable — there is no record of what else was tried and discarded.
03Comparability
Method that holds up when someone checks.
A comparison is only worth as much as the guarantee that every entry in it faced the same conditions.
- Identical data, split and threshold
- The comparison stage fixes the data, the split design and the decision threshold across every estimator it ranks, so a difference in the table is a difference in the model, not the setup.
- An append-only trial ledger
- Every trial recorded stays recorded; there is no update or delete for a trial once it exists, and closing a ledger blocks further appends. A dropped configuration cannot be quietly removed from the count that deflates the winner.
- Combinatorial purged cross-validation available
- Beyond a single chronological split, rolling and expanding walk-forward, CPCV produces many purged, embargoed train/test arrangements from one dataset when the research question calls for it.
- A degradation statistic on every backtest
- In-sample-to-out-of-sample Sharpe retention is graded pass, warn or fail — stated as the number that should be the headline, not the in-sample Sharpe.
04Objections, taken seriously
The questions this audience actually asks.
I already have Python. Why leave the notebook?
You do not have to. The workspace is where the validation and the record live; the notebook can stay where the exploration lives. What changes is that a result now carries how it was made.
Can I bring my own features or models?
Features, yes, through the feature engineering registry across price, statistical, volatility, microstructure, fundamental and cross-sectional families. Bringing a custom script is not available today — Quant Code, a space for custom scripts inside the workspace, is a possible future direction, not a shipped feature.
How is the trial count recorded?
The trial ledger records every configuration run, append-only, and the Deflated Sharpe Ratio and Probability of Backtest Overfitting both consume that count directly rather than a number typed in afterwards.
Does it execute?
No. Nothing in the workspace places or routes an order — the health endpoint reports live order routing as disabled. This is a research and paper-mode workspace.
05In beta, for you
What a researcher gets during the private beta.
- Full access to the 119-method catalogue, the 43-model registry and the four validation split modes
- The trial ledger and Deflated Sharpe / PBO corrections, live on real runs
- PDF, HTML, DOCX and JSON export of the same evidence model
- Direct feedback into the roadmap, including which unverified methods get a reference next
Other audiences, and further reading
Related capabilities
Run your next idea with a record attached.
Beta researchers start from a template, swap in their own data, and keep every trial the ledger sees.