Local Regression Studio v1.0.11

Step 1

Start and load data

CSV

Drop a CSV here or choose a file Recommended: up to 50,000 rows and 50 original columns

Open project Load fitted model Load approved prediction package Use offline mode — local libraries only

Offline mode reloads the app without attempting any CDN request. In either mode, CSV data is never sent to a CDN.

Step 2

Select features and transform the target

Target column Target transformation Feature search

Predictions and uncertainty bounds are converted back to the original target scale.

Data-quality assistant

Automated checks update with the selected target and features. They identify common risks but do not replace scientific or domain review.

Load a CSV to run data-quality checks.

Step 3

Configure preprocessing

Numeric missing values Numeric scaling fitted on training data Categorical missing values Categorical encoding Maximum categories per column Remove one reference category from each categorical featureRecommended for linear and ridge regression to avoid redundant one-hot columns. Usually unnecessary for tree models.

Select a target and at least one feature to estimate the processed feature count.

Step 5

Split data and train the model

Split strategy Random seed Run k-fold cross-validation on training dataFits preprocessing and the model separately in each fold.

Prediction uncertainty

Estimate prediction intervalsThe method depends on the selected model and is labelled in results. Requested interval coverage Bootstrap models for ridge

Choose a split strategy. A split summary will be shown before training.

Compare several baseline models

Single-model training is the default. Open the comparison tools only when you want a quick baseline comparison across multiple model families.

Shared-split rule: quick baseline comparison freezes the current target, selected features, preprocessing, target transformation, uncertainty level, random seed, and exact train/validation/test rows, then trains each selected model once using documented baseline settings.

Show baseline comparison settings

These presets are not the same as full Step 4 tuning. They are quick, documented starting configurations used to compare model families on one frozen split. To tune a model, use Step 4 and train it separately; tuned comparison remains a later workflow.

Model	Baseline settings used in comparison	Speed note
Linear	Ordinary linear regression, no model-specific hyperparameters.	Fast baseline.
Ridge	λ = 1.	Fast; useful with correlated features.
Elastic net	λ = 0.1, L1 ratio = 0.5, coordinate-descent limit = 500.	Moderate; can shrink unhelpful features.
Robust regression	Huber threshold = 1.345, IRLS limit = 80.	Moderate; downweights large residuals.
Decision tree	Depth = 8, minimum leaf = 5, 24 candidate thresholds.	Fast nonlinear baseline.
Random forest	40 trees, depth = 8, minimum leaf = 5, row sample = 0.8, sqrt features.	Moderate; strong general tabular baseline.
Gradient boosting	80 stages, learning rate = 0.05, depth = 3, row sample = 0.8.	Moderate to slow; sequential tree ensemble.
k-nearest neighbours	k = 7, inverse-distance weighting, Euclidean distance.	Fast fit; prediction can be heavier on large data.
Quantile regression	Median with lower/upper quantiles from the requested interval, 1500 iterations.	Moderate; interval behaviour differs from residual intervals.
Gaussian process	RBF kernel, length scale = 1, signal variance = 1, noise = 0.1, stochastic kernel-search iterations = 8; representative subset if needed.	Can be slow; native uncertainty.
Neural network	64 → 32 hidden layers, ReLU, Adam, learning rate = 0.001, dropout = 0.1, early stopping.	Can be slow and iterative.

No comparison batch is running.

Training data fits preprocessing and the model. Validation data selects hyperparameters. Test data is used only for the final independent evaluation.

Step 6

Review, validate, approve, and export

Diagnostic plot settings

Dataset shown in plots Input feature for residual plot Source column for grouped diagnostics

Show uncertainty intervals in plotsThis controls only the displayed uncertainty bands and error bars. It does not change the fitted model or uncertainty metrics.

Actual vs predicted

Actual and predicted vs input feature

Residuals vs predicted

Residual distribution

Residual Q–Q plot

Residuals vs input feature

Residuals by source

Interval width vs prediction

Training and optimisation history

Feature importance

Validation and acceptance

Acceptance criteria apply to the currently active experiment unless you explicitly apply them to all comparable experiments. Blank numeric fields are not evaluated. Acceptance is a user decision, not an automatic certification.

Maximum test RMSE Minimum test R² Minimum interval coverage Maximum interval coverage Maximum group test RMSE Group column

Require no critical data-quality findingsCritical findings are based on the currently selected target and features.

Train or load a model to evaluate acceptance criteria.

Test performance by selected group

Approval and operational release

Record a human decision, document intended and prohibited uses, define the operational input schema, and export an integrity-checked prediction package.

Model selected for approval

Approval decisions apply to this explicit release candidate. Final approval records, history, and approved-package exports are available in Final reports and exports below. The selector defaults to the preferred experiment when one is marked; otherwise it uses the active experiment.

Approval candidate

Train or open a model to choose an approval candidate.

Model name Model developer Model owner Independent reviewer Approver Recorded role Approval status Decision date Next review date

Intended use Prohibited or unsupported uses Known limitations Approval conditions

Decision notes

Operational input schema

Training-derived ranges and category levels are locked. Optional units and descriptions are saved with approved prediction packages.

Train or load a model to record an approval decision.

Approval history

Final reports and exports

Download validation evidence, approval records, operational packages, model artifacts, metrics and plots after completing diagnostics, acceptance review and approval decisions.

Figures to embed in the validation report Actual vs predicted Actual and predicted vs selected feature Residuals vs predicted Residual distribution Q–Q plot Model-comparison chart

Validation and governance reports

Model, project and result exports

The project file does not contain the original CSV. It contains configuration, fitted model, predictions and diagnostic results.

Operational release

The approved prediction package is enabled only when the selected approval candidate has an approved or conditionally approved status.

Step 7

Predict unknown data

Full workspace

No approved prediction package is loaded. Full-workspace users may still predict with the active fitted model.

Upload prediction CSV Plot predictions against input feature Optional source/group column Include original input features in export

Optional measured-target comparison

Show measured target values in the plotMeasured values are displayed only for comparison. They are not used to fit, update or tune the model. Measured target column

Train or load a fitted model, then upload a compatible CSV. A measured target column is optional.

Predictions against selected input feature

Step 8

Monitor operational performance and revalidation

Import a CSV containing previous predictions and later measured outcomes. Monitoring does not retrain or alter the model.

Upload monitoring CSV

Monitoring CSV formats: a minimal file needs measured and predicted target columns. Extended monitoring should also include original input features, package ID, row/sample ID, prediction date, applicability status, and source/regime columns so you can investigate drift and failures.

Upload an operational-results CSV to calculate performance after deployment.

Revalidation triggers

Maximum monitoring RMSE Maximum absolute bias Minimum interval coverage Maximum warning/outside rate Required review date

No monitoring record has been analysed.

Model-change assessment

Compare the active experiment with another saved experiment before replacement or reapproval.

Reference experiment

Save at least two experiments in the project to compare a proposed replacement.