Processing Background Task
0%
Initializing...
v5.7by E|M|E|F Lab
Classification ModeEstimation Mode

1. Data Source & Cleaning

Data: .xlsx, .xls or .csv. A session (.4cl) keeps the data, all settings and the run history.
Observations: 0 Variables: 0

2. Model & Evaluation Settings

Warning: Random Sampling Split provides no protection against data leakage. If the target or any feature is computed over a rolling window (e.g. rolling Sharpe, moving average, RSI, volatility over N periods), time-adjacent observations are nearly identical and may end up split between the training set and the Test Set, artificially inflating the Test Set score (especially for KNN, decision trees, Random Forest). If this applies to your data, prefer Time-Series (Sequential Split).
The Full Report and the Methods Report are in the Analysis Reports tab, where both can be read on screen before printing.

2. Estimation Settings

The dependent variable is the Target Column above and the regressors are the selected features. Nothing is split, scaled or tuned in this mode.

AI Interpretation Your data and AI providers
4Cast Lab does not keep, collect or send any of your data: everything is computed on this computer.

When you click Generate AI Interpretation, the results shown in the app (metrics, coefficient tables and test results, not the raw data file) are sent to the provider you choose below, outside 4Cast Lab. Their handling of that text is governed by the provider's own terms. Do not use this feature with confidential results.

Your API key is used only for that request. It stays in memory and is never written into sessions, reports or files.

3. Probability Cutoff

AUC, log loss and Brier score do not depend on this cutoff; accuracy, sensitivity, specificity, F1, balanced accuracy and cost do.

Bootstrap Confidence Intervals?

Train a model first. The bootstrap resamples the data that the trained model was fitted and evaluated on, so it needs a completed run.
Nothing here runs automatically: the training-set experiments refit the model once per replication, which is the expensive part. Choose the experiments and the number of replications, then press Run Bootstrap.
Changing this re-draws the tables and figures from the replications already computed; it does not re-run the bootstrap. Where the histogram is visibly skewed relative to the fitted normal curve, prefer the percentile interval.
point estimate bootstrap mean 90% limits 95% limits 99% limits red curve: fitted normal N(mean, SD)

Model Used for These Forecasts

Train a model first. The algorithm and the hyperparameters actually used to produce the forecasts will be recorded here.
Estimation mode. Results describe the relationship in the estimation sample, in the original units. They are not evidence of forecast accuracy: for that, send the model to Forecasting mode.
Choose an estimator on the left and click Estimate.
Estimate a model first. Specification, diagnostic and hypothesis tests appear here.
Impulse responses, variance decompositions, long-run effects, volatility and forecasts appear here for the time-series estimators.

Results Table

Each estimated specification can be added as a column with Add to Results Table. The table follows the usual journal layout and exports to LaTeX, Word and Excel.
No specifications yet.

Analysis Reports

The Full Report follows the analysis from start to end: the data and everything done to it, the design, the methods, every table and figure, the diagnostics, the uncertainty of the metrics, the run history and the methodology checklist. The Methods Report is the prose for the Methods section of a paper or thesis. Every report can be read here, printed or saved as PDF, or downloaded as .html.
Train a model, then choose a report above. Its full text and tables appear here and can be read on screen before printing.

Variable Selection?

Choose the methods and press Run variable selection.

Import Report and Missing Values?

Load a dataset to see how it was read.

Time Index and Data Integrity?

Load a dataset to check its time index.

Variable Construction?

Build differences, logs, growth rates, seasonal differences, detrended series and lags from the loaded data. Anything you add appears at once in Target Column (Y) and in Select Features (X) on the left. The original variables are always kept, so you can model levels and transforms side by side.
New variable(s):

Stationarity and Unit Root Tests?

Four tests are run on the target and on every selected predictor, in levels and in first differences. p-values come from MacKinnon's response surfaces, and the lag order and Newey-West bandwidth behind each statistic are printed so the result is reproducible.
The order of integration is a property of the series, so the full sample is the usual choice and is the default. Restrict to the training window if you would rather no hold-out observation informed the specification; the choice is recorded in the Methods Report either way.
Order of integration

Serial Dependence: ACF, PACF, Rolling Moments and CUSUM?

ACF and PACF
Rolling mean and standard deviation

Cross-Correlation with the Target?

Cross-correlation function
4Cast Lab: Analyze / Forecast / Decide. Turning data into smarter decisions.

Machine Learning, Econometrics and Causality

No coding needed. Python when you want it.

4Cast Lab is a free research and teaching laboratory for students, instructors and researchers. Everything runs on your own computer: your data never leaves it.

Forecasting (ML) trains, tunes and compares machine learning models honestly, out of sample, explains what drives them, and tests causality (ML Granger causality, Double ML with refutation checks). Estimation (Econometrics) estimates and tests econometric models: OLS with robust inference, logit and probit, instrumental variables, ARIMA, ARDL with the bounds test, error-correction models, VAR and VECM, and GARCH. AutoML builds a forecasting model in four steps, with an honest verdict in plain words. The Python Workbench runs Python inside the app, with a script of every run that checks each number.

Every result comes with a "How to read this" note and an "In plain words" verdict, and the Methods and Results reports are ready for a thesis or a paper.

EMEF Research Lab
(c) Periklis Gogas
EMEF Lab
Department of Economics
Democritus University of Thrace
For any comments, suggestions, corrections and issues please use the following form:
How to cite. Gogas, P. (2026). 4Cast Lab (Version 5.7) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.22883216
This DOI always resolves to the latest release; each release also has its own version DOI on Zenodo.
Known issues and reporting problems: github.com/PerryGogas/4Cast-Lab/issues
Updates.
The check asks GitHub only for the number of the latest release (at most once a day). No data, file names or results are sent, and nothing is downloaded or installed automatically. Earlier versions stay available on the releases page for reproducing past results.

Data Visualization

Descriptive Statistics & Multicollinearity (VIF)

Select dataset to calculate statistics.

Correlation Matrix Heatmap

Load a dataset and choose a target to see the correlations.

Predictors ranked by correlation with the target

Descriptive: one predictor at a time, no model, no cross-validation.

Optimized Hyperparameters

Run training to execute grid search.

Class Scatter: Two Predictors?

ROC Curve / Regression Fit

Multi-Criterion Feature Ranking

Run pipeline to evaluate importance scores.

Partial Dependence Plot (PDP)

Shapley Values (SHAP Summary)?

Learning Curve (Bias vs. Variance)?

Residual Analysis (Errors vs Fitted)

Model Performance Metrics

Run evaluation to see results.

Confusion Matrices

Training Set

Test Set

Predictions Output Tables

Training Set Predictions

Test Set Predictions