FastData Help
Context-sensitive help entries for FastData application
Project Links
Introduction
Using This Help
Start here to understand how the help content is organized.
This help guide is organized around the workflows and concepts you use in FastData.
- Basics explains the core data and modeling concepts used across the app.
- Features groups the main workflow areas and the controls available in each one.
Use the navigation on the left to jump directly to the part of the application you are working with.
Basics
Data Concepts
Files and Sheets (Imports)
Files/sheets are stored as imports; for users these are effectively the same thing.
The first concept is the import. In practice, users select files and sheets, and each selected file/sheet becomes an import.
Technically, an import is represented as a database table. From the user perspective, it is fine to think of file/sheet = import.
Typical import sources:
- CSV files
- Excel files and individual sheets
- Other supported tabular sources
Imports are the atomic source units that later get grouped into datasets.
Datasets
A dataset is a collection of one or more imports used together for analysis.
A dataset groups related imports into one analysis scope.
One dataset can contain many imports (for example, multiple files/sheets from the same process).
The active dataset drives most work in the app: filtering, preprocessing, visualization, and modeling.
Use datasets to separate contexts like production lines, time periods, or raw vs cleaned data.
Systems
A system contains multiple datasets and its own feature set.
A system is a higher-level container above datasets.
Each system can hold multiple datasets, and each system owns its own feature set.
This means feature definitions are managed in system context, even when data is split across multiple datasets in that same system.
Use systems to separate fundamentally different processes, plants, machines, or product families.
Features
Features are columns used for analysis and modeling inside a system.
Features are the measurable variables (columns) used as model inputs or analysis dimensions.
Because systems own feature sets, feature metadata is managed consistently across datasets within the same system.
For each feature, users should understand at least:
- Source: where the measurement comes from (sensor, calculated field, instrument, etc.)
- Unit: physical/logical unit (for example C, bar, kg/h, %)
- Type: value type like how it was measured (statistical, raw, etc.)
Clear feature metadata improves filtering, preprocessing, model quality, and interpretation.
Tags
Tags label features for organization, filtering, and workflow clarity.
Tags are lightweight labels you assign to features (and optionally other entities) to keep large models understandable.
Typical tag examples: target, input, quality, temperature, critical, lab, calculated.
Use tags to:
- Group related features quickly
- Filter feature lists for modeling or monitoring
- Keep naming simple while still preserving context
Tags are metadata only. They do not change raw measured values.
Data Flow
Canonical structure: files/sheets -> datasets -> systems -> features -> tags.
The conceptual model is:
- Files/Sheets are ingested as imports (database tables)
- Datasets group one or many imports
- Systems group multiple datasets and own feature definitions
- Features define the usable variables with source/unit/type metadata
- Tags label features for organization and filtering
Operationally, you import files first, assign/group them into datasets, work inside a system context, then manage features and tags for efficient analysis.
Target Variable
The variable you want to predict or model.
The Target Variable (also called the dependent variable or label) is what your model is trying to predict.
Examples:
- In a sales forecasting model, the target is sales amount
- In a classification problem, the target is the category or class
- In regression, the target is a continuous numeric value
The target variable should be clearly defined before beginning model training.
Imported Feature
A feature column with source, unit, type, and optional tags.
This is a feature column available in the current system/dataset context.
Check and maintain feature metadata:
- Source - Origin of this feature value
- Unit - Unit of the feature value (e.g., %, $, °C)
- Type - How the feature value was measured (e.g., raw, normalized, statistical)
- Tags - Labels for grouping and filtering
To use this feature effectively:
- Check for missing values and outliers
- Understand its distribution and data type
- Consider if preprocessing (scaling, encoding) is needed
- Evaluate its correlation with the target variable
You can view detailed statistics and visualizations for this feature in the data exploration tab.
Modeling
Model Selection
Choose the best algorithm for your prediction task.
Model Selection is the process of choosing the most appropriate machine learning algorithm for your specific problem.
Key considerations:
- Problem Type - Regression, classification, clustering, or forecasting?
- Data Size - Some algorithms work better with small vs. large datasets
- Interpretability - Do you need to explain predictions? (Linear models > Neural networks)
- Performance Metrics - Use cross-validation to compare models objectively
- Training Time - Balance accuracy vs. computational cost
FastData lets you run and compare multiple models so you can choose the best fit based on your metrics and constraints.
Features
Data
Import data, manage databases, and define the active selection scope.
The Data tab is the main entry point for preparing data in FastData.
- Import files into the project database (Import Files…).
- Create, open, and save database files from the sidebar actions.
- Use Data selection to choose systems, datasets, imports, tags, filters, and preprocessing.
- Preview the resulting table to confirm values and columns before analysis.
Selections configured here are reused by downstream tabs through the same data-selection model.
Filters
Data Filters
Apply conditions to subset your dataset.
Filters allow you to select specific rows from your dataset based on conditions.
Use filters to:
- Focus on specific time periods
- Exclude outliers or invalid data
- Analyze specific categories or groups
- Compare subsets of data
The selected filter set is applied to the dataframe used by downstream analysis and visualizations.
Systems and datasets
Scope results by system and dataset in one place.
Use these selectors together to narrow the data to specific systems and the datasets they contain.
- Check multiple systems or datasets to compare them side by side.
- Leave either list empty to include every option for that category.
- Combine both filters to isolate behavior for chosen datasets within selected systems.
Date range
Limit data between a start and end timestamp.
Pick a starting and ending date/time to bound the records that are loaded, previewed, and charted.
- Set only the start to include everything after that point.
- Set only the end to include everything up to that timestamp.
- Leave both empty to include the full available history.
Months and groups
Filter by calendar months and predefined groups.
Select calendar months to focus on seasonal patterns, and choose database groups to restrict which entities are processed.
- Pick multiple months to compare seasons such as winter vs. summer.
- Combine month and group filters to study specific cohorts during chosen periods.
- Leave a list empty to include all options for that filter.
Imports and tags
Filter by selected imports and feature tags together.
Use Imports to restrict data to specific ingestion events in the currently selected dataset.
- Each import entry corresponds to a file ingestion (file name + import timestamp).
- Select one or many imports to compare runs or isolate a single load.
- Leave imports empty to include all imports in scope.
Use Tags to filter which features are available/selected for analysis.
- Tags are feature metadata labels (for example, domain or equipment class).
- Select multiple tags to include features that match those labels.
- Combine imports + tags to narrow both rows (data provenance) and columns (feature set).
Import Options
System name
Label the imported data with a system name.
Choose the system that owns the data you are importing.
- Existing systems appear in the list; you can type a new name to create one.
- The system is used to organize datasets and imports in the database.
Dataset name
Attach the data to a specific dataset.
Pick a dataset that belongs to the selected system.
- Use the dropdown to pick an existing dataset or type a new one.
- <Use sheet name> uses the Excel sheet name as the dataset value.
Header row amount
Number of header rows in Excel files.
Define how many rows at the top of the sheet are header rows.
- Use 0 if the sheet has no header row.
- This setting is only used for Excel-style files.
Header delimiter
Split multi-part headers by a delimiter.
Split header text into multiple parts using a delimiter (for example "_" or " | ").
- Use this when the header contains concatenated fields (e.g., Temp_Outdoor_C).
- Leave empty to keep the header text as-is.
Base name row
Choose which header row contains the base feature name.
Select the header row that contains the base feature name (e.g., Temperature).
- Set to None if the file does not contain that row.
- Combine with the header delimiter to split values into parts.
Source row
Select the header row for source/series labels.
Use this when a header row identifies the source or series for a feature.
- Examples: phase identifiers or channel names.
- Leave unset if the file does not include source rows.
Unit row
Select the header row containing units.
Point to the header row that contains measurement units (e.g., °C, kW).
- Units are stored with the feature metadata.
- Leave unset if units are already embedded in names.
Type row
Select the header row for qualifiers or annotations.
Qualifiers capture extra descriptors like min/max, status, or quality.
- Use when headers include additional descriptors beyond base name and unit.
- Leave unset if not applicable.
Force meta columns
Treat specific columns as metadata (not measurements).
Comma-separated list of column names to force into metadata.
- Use this to prevent ID or label columns from becoming features.
- Matching is case-insensitive.
Ignore column prefixes
Skip columns that start with listed prefixes.
Comma-separated prefixes to ignore when importing columns.
- Example: Time, AE_ will ignore columns starting with those values.
- Use this to exclude helper columns from imports.
Date column
Choose which column contains timestamps.
Select the column that holds datetime values.
- Leave as auto to let the importer detect a timestamp column.
- Set explicitly if the detection picks the wrong column.
Assume day-first dates
Interpret 03/04 as 3 April instead of 4 March.
Enable this if your date strings use day/month order.
- Useful for European date formats.
- Disable for US month-first formats.
Dot time formatting
Parse 9.00 as 09:00.
Convert dot-separated times to colon-separated times while parsing.
- Example: 13.47 becomes 13:47:00.
- Disable if dots are not used as time separators.
Datetime formats
Provide explicit datetime formats (optional).
Comma-separated list of strftime-compatible formats.
- Use auto to let the importer guess formats.
- Supplying formats can speed up parsing and avoid ambiguities.
CSV delimiter
Character used to separate CSV columns.
Set the delimiter between columns (for example , or ;).
- Leave as auto to let the importer detect the delimiter.
- Overrides the quick guess if you enter a value.
CSV decimal
Decimal separator for numeric values.
Specify the decimal character for numeric values.
- Common values are . and ,.
- Leave as auto to use the default parser behavior.
CSV encoding
Text encoding for CSV files.
Provide an encoding name (for example utf-8 or latin-1).
- Leave as auto to try common encodings automatically.
- Set explicitly when you see garbled characters.
Use DuckDB CSV import
Load large CSVs with DuckDB's fast parser.
Enable DuckDB's CSV import pipeline for large files.
- Recommended for very large CSVs (100MB+).
- Disable if you need strict pandas-style parsing behavior.
Preprocessing
Preprocessing
Prepare selected data with the built-in preprocessing controls.
Preprocessing in FastData is configured from the shared Data selection area used across tabs.
The app currently provides these preprocessing options:
- Timestep: keep native spacing, use a preset interval, or enter custom seconds.
- Moving average: optional smoothing window (preset or custom seconds).
- Fill empty: fill missing timestamps with none, zero, previous, or next value.
- Aggregation: combine overlapping points using avg, min, max, first, last, or median.
These controls shape the dataframe used by Statistics, Charts, SOM, Regression, and other analysis flows.
Timestep
Resample data to a fixed interval or leave it on auto.
Choose the interval used to resample incoming measurements.
- auto keeps the native spacing of the data.
- Preset values such as 1 minute or 1 hour resample the series to an even grid.
- Enter a custom number of seconds to match an exact cadence.
Resampling can make downstream statistics and charts easier to compare.
Moving average
Smooth measurements with an optional rolling window.
Applies a rolling mean over the selected window to reduce noise.
- none leaves the data untouched.
- Preset windows (e.g., 5 minutes) smooth short-term spikes.
- Type a custom window in seconds for finer control.
Use smoothing when charts or models are sensitive to rapid fluctuations.
Fill empty
Choose how to handle missing timestamps after resampling.
Determines how gaps created during resampling are filled.
- none leaves gaps as-is.
- zero inserts 0 for missing values.
- prev carries the previous value forward.
- next uses the next known value.
Forward or backward filling is useful when signals change slowly and occasional gaps appear.
Aggregation
Summarize multiple points that fall within a timestep.
When multiple measurements land in the same resampled bucket, aggregation decides how they are combined.
- avg (default) computes the mean value.
- min/max keep the extremes.
- first/last preserve ordering-sensitive data.
- median is robust against outliers.
Pick the method that best matches how you would summarize overlapping readings.
Selections
Create reusable feature selections and filter presets.
The Selections tab lets you curate which features to keep for downstream tasks.
- Toggle columns on or off to build a focused feature set.
- Right-click a saved feature to convert its measurement values into database groups.
- Use Reload features to restore the table from the database and discard unsaved edits.
- Choose a filter scope per feature to apply value limits globally, by system, dataset, import, or only to that feature.
- Store multiple presets and quickly switch between them for experimentation.
- Export or import selection databases to share with teammates.
Adjusting selections here keeps your preprocessing and modeling steps consistent.
Statistics
Review derived statistics and prepared measurements.
The Statistics tab focuses on validating the numbers produced from your data.
- Preview aggregated measurements and sanity-check values before saving them.
- See how the shared data-selection filters and preprocessing settings shape the resulting metrics.
- Confirm column names, units, and sampling windows prior to exporting.
This tab focuses on statistical outputs, while preprocessing is configured in the shared Data selection widget used by this tab.
Statistics
Statistics actions
Run the computation and store results when you are happy with them.
Use these buttons to generate and persist the statistics previewed in this tab.
- Gather statistics executes the selected statistics with the current filters, mode, and preprocessing settings.
- Save statistics writes the last computed preview into the database so it can be reused in other tabs.
- Run again any time you tweak filters, periods, or selected statistics to refresh the preview.
Statistics to compute
Pick the summary measures calculated for each time bucket or group.
Select one or more statistics to include in the output. Each choice adds a column to the preview and saved results.
- Average (avg/mean): mean value of the samples.
- Minimum / Maximum: smallest or largest observation.
- Median: middle value that is robust to outliers.
- Std Dev: spread of the samples around the mean.
- Sum: total of all values in the bucket.
- Count: number of rows contributing to the statistic.
- Interquartile Range (IQR): difference between the 75th and 25th percentiles.
- Outliers (3σ %): percentage of points that land three standard deviations from the mean.
- Max derivative: largest absolute change between consecutive samples.
Pick only what you need to keep previews fast; you can always rerun with more metrics.
Aggregation mode
Choose whether to aggregate over time or by a specific column.
- Time based groups measurements into regular periods using the Statistics period setting.
- Group by column ignores time buckets and instead aggregates by the values of the chosen Group column.
Switching modes changes how the preview table and chart summarize your data.
Group column
Pick the categorical column that defines each group in column mode.
Available options come from the dataset and only apply when Group by column is selected.
- Choose identifiers like system, dataset, or any other categorical field.
- Each distinct value becomes its own row or bar in the preview.
- Leave the dropdown empty in time-based mode—the control is disabled automatically.
Statistics period
Define the window used for time-based aggregation.
Time-based mode buckets records into regular windows before applying your chosen statistics.
- hourly, daily, weekly, monthly: convenient presets that translate to fixed second intervals.
- custom: enter a number of seconds in Custom period (s) to match your own cadence.
- Pick a shorter window for fine-grained trends or a longer one for smoother, aggregate views.
Separate timeframes
Choose whether group-kind statistics are split by each saved timeframe segment.
This setting applies when using Group by column with a database group kind (for example group:som_cluster).
- Enabled: each saved timeframe segment is treated as its own group bucket.
- Disabled: all rows with the same group label are merged into one bucket, regardless of timeframe segments.
Charts
Visualize your dataset with quick plots.
The Charts tab provides fast visual feedback on your data.
- Create common plots such as histograms, scatter plots, and correlations.
- Use selections and filters to focus charts on specific subsets.
- Pick a target feature and run Find feature correlations to auto-build: a top-10 correlation bar chart.
- Compare relationships between variables before modeling.
Use charts to spot trends, outliers, and data quality issues early.
SOM
SOM
Explore data structure with Self-Organizing Maps.
The SOM tab visualizes high-dimensional data on a 2D grid.
- Train self-organizing maps to cluster similar observations.
- Inspect component planes and distance maps to spot patterns.
- Experiment with map sizes and training parameters.
Use this view for exploratory analysis and for identifying interesting cohorts.
Self-Organizing Maps (SOM)
Unsupervised neural network for data visualization and clustering.
Self-Organizing Maps (also called Kohonen maps) are a type of artificial neural network trained using unsupervised learning to produce a low-dimensional representation of the input space.
Key features:
- Dimensionality Reduction - Projects high-dimensional data onto a 2D grid
- Topology Preservation - Similar data points are mapped to nearby locations
- Clustering - Automatically groups similar observations
- Anomaly Detection - Outliers appear in distant or sparse regions
SOMs are particularly useful for exploratory data analysis and pattern recognition.
FastData uses the MiniSom library for efficient SOM computation.
Clustering
Feature clustering model
Pick the algorithm used to group feature planes.
Choose how features are grouped based on their component planes.
- K-Means / Mini-Batch K-Means – fast centroids, works well as a default.
- Agglomerative – hierarchical merges; good when you expect nested structure.
- Spectral – uses graph cuts; handles complex shapes but needs a few more samples.
All options support automatic K search when allowed by the method.
Max K & Clusters
Control the cluster search range or pin a fixed number.
These two fields work together:
- Max K sets the upper bound for automatic K search (min is 2).
- Clusters overrides auto-search when set; 0 or "Auto" falls back to scoring.
- If the dataset has few samples, the system trims the range to stay valid.
Scoring metric
Metric used to pick the best K during auto-search.
Scores compare candidate clusterings:
- Silhouette – balances cohesion and separation; good general-purpose metric.
- Calinski-Harabasz – favors compact, well-separated clusters; fast to compute.
- Davies-Bouldin – lower is better; sensitive to overlapping clusters.
Only used when Clusters is set to Auto.
Cluster features
Group SOM component planes into feature clusters.
Runs clustering on the trained SOM component planes.
- Requires a trained SOM model; train first if the button is disabled.
- Outputs cluster labels you can inspect on the feature map tab.
- Use the scoring metric to automatically select K when Clusters is Auto.
Auto cluster features
Automatically run feature clustering after SOM training.
When enabled, feature clustering starts automatically each time a SOM model finishes training.
- Uses the current Feature clustering settings (model, Max K, Clusters, Score).
- Runs in the background, the same as clicking Cluster features manually.
- Disable this if you only want to train SOM maps without updating feature clusters.
Neuron clustering model
Pick the algorithm used to group SOM neurons.
The same clustering algorithms are available for neurons as for features:
- K-Means / Mini-Batch K-Means – fast, dependable defaults.
- Agglomerative – builds a hierarchy of regions on the map.
- Spectral – can separate winding or non-convex neuron regions.
Choose the method that matches how you expect neurons to organize across the grid.
Max K & Clusters (neurons)
Set the range or fixed count for neuron groups.
Configure how many neuron groups to consider:
- Max K caps the auto-search; higher values explore finer partitions.
- Clusters locks in an exact count when set to a value > 0.
- If the SOM grid is small, the available K values are trimmed to remain valid.
Scoring metric (neurons)
Metric used when auto-selecting neuron clusters.
Uses the same metrics as feature clustering:
- Silhouette – balanced, works well on most maps.
- Calinski-Harabasz – rewards tight neuron groups.
- Davies-Bouldin – lower is better; highlights overlapping groups.
Cluster neurons
Partition the SOM grid into neuron clusters.
Runs clustering on the neurons themselves, using their codebook vectors.
- Requires a trained SOM model before it can run.
- Results can be visualized on the maps to highlight regional patterns.
- Use Max K, Clusters, and Score to tune how fine-grained the regions are.
Auto cluster timeline
Automatically run neuron clustering after SOM training.
When enabled, neuron clustering starts automatically after each SOM training run.
- Uses the current Neuron clustering settings (model, Max K, Clusters, Score).
- Keeps timeline cluster mode and cluster map ready without a separate manual step.
- Disable this if you want manual control over when timeline clusters are recalculated.
Controls
Cluster Timeline
Shows BMU/cluster states over time, with optional selected-feature overlays.
This chart plots timeline layers selected in Display.
- BMU shows the raw winning neuron index over time.
- Neuron clusters shows clustered neuron IDs after clustering.
- Selected features overlays scaled feature traces on top of the timeline.
- Use this to spot regime shifts and align feature changes with BMU/cluster transitions.
Timeline Display Layers
Select one or more layers: BMU, neuron clusters, and selected features.
Use the Display multi-select control to choose timeline layers:
- BMU: shows the raw winning neuron index over time.
- Neuron clusters: shows clustered neuron IDs after clustering.
- Selected features: overlays feature traces for currently selected features.
Data Table
Row-level BMU assignments used by the timeline.
This table lists BMU assignments for each data row.
- Select rows to highlight their neurons on the cluster map.
- Use it to inspect specific time ranges in detail.
Cluster Map
Neuron grid colored by cluster assignment.
The cluster map colors each neuron by its cluster ID.
- Use it to see spatial structure in the clustering.
- Click neurons to inspect specific cluster groups.
Save as timeframes
Store contiguous cluster runs as start/end ranges instead of single points.
When saving timeline clusters, this option controls how group assignments are stored.
- Enabled: consecutive rows with the same cluster label are merged into one timeframe range.
- Disabled: each timeline row is saved as an individual timestamp assignment.
- Timeframe mode improves readability in statistics and group-based analysis by preserving segment boundaries.
Hyperparameters
Map width
Horizontal size of the SOM grid (auto if left blank).
Map width controls how many neurons the map has along the X-axis.
- Leave the field empty to let FastData pick a balanced width based on your dataset size (roughly
sqrt(5 * sqrt(N))). - Enter an integer to force a specific width. Values below 2 are clamped to a minimum grid size.
- Wider maps expose more local structure but require more samples to train well.
Map height
Vertical size of the SOM grid (auto if left blank).
Map height controls how many neurons the map has along the Y-axis.
- Leave it empty to reuse the automatic heuristic used for width.
- Specify a value to make rectangular maps (e.g., wider than tall) when you expect directional structure.
- Both width and height must be positive; the UI enforces a minimum of 2.
Sigma
Initial neighborhood radius for SOM training.
Sigma sets how far the learning influence of a winning neuron spreads across the grid at the start of training.
- Larger values encourage smoother, more global organization early on.
- Smaller values keep neighborhoods tight, emphasizing local differences.
- The value decays during training; 6.0 is a sensible default for most datasets.
Learning rate
Step size used when updating SOM weights.
Learning rate controls how quickly the map adapts to each sample.
- Higher values (e.g., 0.5) converge faster but can overshoot fine structure.
- Lower values make training steadier at the cost of more iterations.
- Learning rate decays over epochs to stabilize the final map.
Epochs
How many training passes to run (defaults to 100).
Epochs determines how many times the algorithm sweeps through your data.
- Leave at
0or empty to use the default of 100 epochs. - More epochs improve convergence but increase training time.
- If training stalls, try a few extra epochs together with a smaller learning rate.
Normalisation
How features are scaled before training.
Scaling features keeps each variable on a comparable range.
- Z-score (default) standardizes to zero mean and unit variance—best general choice.
- Min-max rescales each column to [0, 1] to preserve relative spacing.
- None skips scaling; use only when your inputs are already normalized.
Training mode
Choose between batch and random SOM updates.
Pick the strategy used to present samples during training.
- Batch (default) updates the entire map per epoch for stable, reproducible results.
- Random feeds one sample at a time in random order, which can discover sharp local structures.
- If results look noisy, switch back to batch mode or increase sigma.
Regression
Regression
Build and evaluate regression models.
The Regression tab guides you through training models for continuous targets.
- Select algorithms, hyperparameters, and cross-validation settings.
- Train models on the current selection and preprocessing pipeline.
- Review metrics and compare runs to pick the best performer.
Use this tab when predicting numeric outcomes such as prices, temperatures, or measurements.
Regression Analysis
Build predictive models to estimate continuous values.
Regression is a statistical method for modeling the relationship between a dependent variable and one or more independent variables (features).
FastData supports multiple regression algorithms including:
- Linear Regression - Simple, interpretable baseline
- Ridge/Lasso - Regularized linear models to prevent overfitting
- Random Forest - Ensemble method for complex non-linear relationships
- Gradient Boosting - Advanced ensemble technique for best performance
Use regression when your target variable is continuous (e.g., price, temperature, sales volume).
Cross Validation
Cross-validation strategy
Pick how training/validation folds are built.
Cross-validation helps estimate how well a model generalises. Choose a strategy that matches your data.
- No cross-validation: train once on the full training set.
- K-Fold: split rows into k equal parts and rotate the holdout fold.
- Stratified K-Fold: like K-Fold but keeps the class/bin distribution of the chosen stratify feature.
- Time series split: preserves order by using earlier rows for training and later rows for validation.
- Group K-Fold: keep all rows from the same group in the same fold to prevent leakage.
Time-series splits build expanding/rolling windows in chronological order, ignore shuffling, and can use a time gap to reduce leakage.
Folds
Set how many slices to use when performing K-Fold style validation.
The number of folds controls how many train/validation rotations are executed.
- Common choices are 5 or 10; higher values increase runtime but give smoother estimates.
- For very small datasets keep folds low so each validation set still contains enough samples.
- Time series splits always step forward through the ordered data regardless of the fold count.
Shuffle folds
Randomly reshuffle rows before forming non-time-series folds.
Enable shuffling to randomise the order of records before K-Fold or Stratified K-Fold splits.
- Recommended when data has any temporal or grouped ordering that could bias folds.
- Automatically disabled for time series validation to preserve chronological order.
- Pair with a fixed random state in model parameters when you need repeatable runs.
Time gap
Skip recent observations between training and validation windows.
When using the Time series split strategy, the gap adds a buffer between the end of the training window and the start of the validation window.
- Use a positive gap to avoid leakage when measurements have autocorrelation.
- Set to 0 to use adjacent windows.
- Time series splits never shuffle; they always move forward in time.
- This setting is ignored for non-time-series strategies.
Stratify by
Balance folds by matching the distribution of a feature.
Select a categorical feature or the target to stratify K-Fold splits.
- Stratification keeps each fold representative when the target has few unique values or strong grouping.
- Works with Stratified K-Fold and test split stratification; ignored for time series validation.
- Leave empty to let the system choose the most suitable option or disable stratification.
Group
Choose the group label used for Group K-Fold.
Select a group kind so all rows from the same group stay in the same fold.
- Use this when data points are clustered by equipment, batch, site, or other group labels.
- If the selected group has too few unique labels, the run falls back to K-Fold.
- Only used with Group K-Fold.
Hyperparameters
Fit intercept
Include an intercept term in the linear regression model.
When enabled, the model learns a bias term in addition to feature weights.
- Disable only when your features are already centered around zero.
- The intercept shifts predictions up or down without changing slopes.
Positive coefficients
Constrain coefficients to be non-negative.
Forces all learned weights to be positive, which can help with interpretability.
- Use when negative contributions are not meaningful for your data.
- Constraint can reduce flexibility and slightly slow down fitting.
Alpha (ridge)
Regularization strength for ridge regression.
Higher values apply stronger L2 penalty and reduce coefficient magnitude.
- Start small and increase if the model overfits.
- Very large values can underfit by shrinking weights too much.
Ridge solver
Numerical method used to fit ridge regression.
Select a solver to match dataset size and stability requirements.
- auto chooses a sensible default.
- svd and cholesky are good for dense data.
- lsqr, sparse_cg, sag, and saga handle large datasets.
- lbfgs supports constrained or sparse scenarios.
- sag and saga typically benefit from standardized features.
Random state (ridge)
Seed for solver randomness where applicable.
Set a fixed value to make the ridge solution reproducible.
Only affects solvers that rely on stochastic optimization.
Alpha (lasso)
Regularization strength for lasso regression.
Higher values apply stronger L1 penalty and promote sparsity.
Too much regularization can drive useful coefficients to zero.
Max iterations (lasso)
Maximum optimizer steps before stopping.
Increase if you see convergence warnings or unstable results.
Larger datasets or high regularization may require more iterations.
Random state (lasso)
Seed for solver randomness.
Fix this value to make lasso results repeatable.
Applies when the optimizer uses random coordinate selection.
Alpha (elastic net)
Regularization strength for elastic net.
Controls the combined L1/L2 penalty magnitude.
Increase to shrink coefficients more aggressively.
L1 ratio
Balance between L1 and L2 penalties.
0.0 is pure ridge (L2), 1.0 is pure lasso (L1).
Intermediate values blend sparsity with coefficient shrinkage.
Max iterations (elastic net)
Maximum optimizer steps before stopping.
Increase if convergence is slow or unstable.
Higher values can improve accuracy on large feature sets.
Random state (elastic net)
Seed for solver randomness.
Fix this value to reproduce elastic net results.
Only relevant when the optimizer is stochastic.
Polynomial degree
Degree of polynomial features.
Higher degrees allow more complex curves but can overfit.
Degrees 2-3 are common starting points for nonlinear trends.
Fit intercept (polynomial)
Include a bias term in the polynomial model.
Enable unless your features are centered around zero.
Disable if you already include a bias feature in preprocessing.
Number of trees (random forest)
How many trees to build in the forest.
More trees improve stability but increase runtime.
Accuracy gains diminish once the forest is large enough.
Max depth (random forest)
Limit the depth of each tree.
Use None to expand until all leaves are pure or minimal.
Shallower trees reduce variance but may underfit.
Min samples split (random forest)
Minimum samples required to split a node.
Higher values reduce overfitting.
Set as a count or fraction of the training samples.
Min samples leaf (random forest)
Minimum samples required in a leaf node.
Higher values smooth the model and reduce variance.
Set as a count or fraction for larger datasets.
Random state (random forest)
Seed for the bootstrap and feature selection.
Fix this value for repeatable forests.
Helps compare runs when tuning other parameters.
Number of trees (extra trees)
How many extra trees to build.
More trees improve stability but increase runtime.
Extra Trees are more randomized, so more estimators help.
Max depth (extra trees)
Limit the depth of each tree.
Use None to expand until all leaves are pure or minimal.
Lower depths reduce variance and improve generalization.
Min samples split (extra trees)
Minimum samples required to split a node.
Higher values reduce overfitting.
Set as a count or fraction of the training samples.
Min samples leaf (extra trees)
Minimum samples required in a leaf node.
Higher values smooth the model and reduce variance.
Larger leaves can help with noisy measurements.
Random state (extra trees)
Seed for randomized splits.
Fix this value for repeatable results.
Controls the randomness of feature and split selection.
Boosting stages
Number of boosting stages to perform.
More stages can improve accuracy but raise risk of overfitting.
Pair with a smaller learning rate for smoother training.
Learning rate (gradient boosting)
Shrinkage applied to each boosting step.
Smaller values require more estimators but can generalize better.
Larger values converge faster but risk overshooting.
Max depth (gradient boosting)
Depth of individual regression trees.
Shallow trees reduce variance but may underfit.
Depth 2-4 is common for stable boosting models.
Min samples split (gradient boosting)
Minimum samples required to split a node.
Higher values reduce model variance.
Set as a count or fraction of the training samples.
Min samples leaf (gradient boosting)
Minimum samples required in a leaf node.
Higher values improve generalization on noisy data.
Larger leaves produce smoother predictions.
Random state (gradient boosting)
Seed for the boosting process.
Fix this value for repeatable results.
Ensures the same subsampling order where applicable.
Number of estimators (AdaBoost)
Number of weak learners to combine.
More estimators can improve accuracy but increase runtime.
Too many stages can overfit noisy data.
Learning rate (AdaBoost)
Contribution of each weak learner.
Lower values require more estimators to reach the same performance.
Higher values can overemphasize errors and reduce stability.
Random state (AdaBoost)
Seed for boosting randomness.
Fix this value for repeatable results.
Helps compare parameter tweaks consistently.
Kernel (SVR)
Kernel function used by support vector regression.
Choose the kernel that best matches the data shape.
- rbf for nonlinear relationships.
- linear for linear trends.
- poly for polynomial curves.
- sigmoid for neural-like boundaries.
- Scale features when using rbf, poly, or sigmoid kernels.
Regularization (C)
Penalty for errors in SVR.
Higher values fit the training data more closely.
Too high can overfit; too low can underfit.
Epsilon (SVR)
Margin of tolerance in the loss function.
Higher values ignore small errors and create a smoother fit.
Larger epsilon usually yields fewer support vectors.
Number of neighbors (KNN)
How many neighbors to average for predictions.
Lower values fit locally; higher values smooth predictions.
Odd numbers can reduce tie votes in classification-like data.
Weights (KNN)
Weighting strategy for neighbors.
- uniform treats all neighbors equally.
- distance gives closer neighbors more influence.
- Distance weighting benefits from scaled features.
Algorithm (KNN)
Search algorithm for nearest neighbors.
- auto selects based on data size.
- ball_tree and kd_tree are optimized for structured data.
- brute checks all points directly.
- High-dimensional data often defaults to brute force.
Max depth (decision tree)
Limit the depth of the tree.
Use None to expand until all leaves are pure or minimal.
Shallower trees are easier to interpret but may underfit.
Min samples split (decision tree)
Minimum samples required to split a node.
Higher values reduce overfitting.
Set as a count or fraction of the training samples.
Min samples leaf (decision tree)
Minimum samples required in a leaf node.
Higher values smooth predictions and improve generalization.
Use larger leaves for noisy or sparse datasets.
Random state (decision tree)
Seed for randomized splits.
Fix this value for repeatable trees.
Useful when comparing depth or split settings.
Layers (MLP)
Sizes of the hidden layers for the neural network.
Enter comma-separated sizes, e.g. 64,32,16 for three layers.
More layers and neurons increase capacity but can overfit.
Activation (MLP)
Activation function used in hidden layers.
relu is a strong default; tanh can help with bounded data.
Solver (MLP)
Optimizer used to train the network.
adam is robust for most datasets.
lbfgs can converge faster on smaller datasets.
Alpha (MLP)
L2 regularization strength for the network.
Higher values apply stronger weight decay.
Learning rate schedule (MLP)
How the learning rate evolves during training.
constant uses a fixed rate.
adaptive reduces the rate when progress stalls.
Max iterations (MLP)
Maximum training epochs.
Increase if the model stops before converging.
Random state (MLP)
Seed for weight initialization.
Fix this value for repeatable runs.
Variance threshold
Remove features with variance below this threshold.
Use higher values to drop low-variance features.
A threshold of 0 removes only constant features.
Number of features (Select K Best)
Pick the top K features by score.
Choose All to keep every feature.
Lower values enforce more aggressive feature selection.
Number of features (Mutual Info)
Pick the top K features by mutual information.
Choose All to keep every feature.
Use smaller K to focus on the strongest nonlinear signals.
Random state (Mutual Info)
Seed for mutual information estimation.
Fix this value for repeatable scores.
Important when comparing ranking stability.
Importance threshold (Random Forest)
Minimum importance required to keep a feature.
Use median, mean, or a numeric value.
Features below the threshold are removed from the dataset.
Number of trees (RF importance)
How many trees to build for importance scores.
More trees give more stable importance estimates.
Smaller values run faster but increase variance.
Random state (RF importance)
Seed for the importance estimator.
Fix this value for repeatable importances.
Useful when comparing thresholds.
Importance threshold (Extra Trees)
Minimum importance required to keep a feature.
Use median, mean, or a numeric value.
Higher thresholds keep only the most influential features.
Number of trees (Extra Trees importance)
How many trees to build for importance scores.
More trees give more stable importance estimates.
Extra Trees are noisy, so more estimators help stability.
Random state (Extra Trees importance)
Seed for the importance estimator.
Fix this value for repeatable importances.
Helps compare threshold values consistently.
Importance threshold (Gradient Boosting)
Minimum importance required to keep a feature.
Use median, mean, or a numeric value.
Higher values prune more aggressive feature sets.
Number of estimators (GB importance)
How many estimators to build for importance scores.
More estimators give more stable importance estimates.
Pair with smaller learning rates for stability.
Random state (GB importance)
Seed for the importance estimator.
Fix this value for repeatable importances.
Use when comparing thresholds across runs.
Features to select (RFE)
Target number of features to keep.
Use None to keep half the features by default.
Smaller targets yield more aggressive reduction.
Features removed per step (RFE)
How many features to eliminate each iteration.
Smaller steps are more precise but take longer.
Use larger steps for faster but coarser selection.
Random state (RFE)
Seed for randomized estimator behavior.
Fix this value for repeatable selection.
Helps verify if rankings are stable across runs.
PCA components
How many principal components to keep.
Use None to keep all components.
Lower values reduce dimensionality more aggressively.
PCA solver
Algorithm used to compute principal components.
auto picks a reasonable solver based on data shape.
full uses a deterministic full SVD.
arpack computes a truncated decomposition.
randomized is faster for large datasets but stochastic.
Whiten components
Scale components to unit variance.
Whitening can help some models but may amplify noise.
Keep it off unless you have a specific reason to enable it.
Random state (PCA)
Seed for randomized PCA behavior.
Fix this value when using the randomized solver to reproduce results.
Components (PLSRegression)
Number of latent variables to keep.
Higher values capture more signal but increase model complexity.
Keep this below the effective rank of your input data.
Scale data (PLSRegression)
Standardize X and y inside the PLS step.
Enable in most cases unless inputs are already consistently scaled.
Max iterations (PLSRegression)
Maximum iterations for the NIPALS solver.
Increase if convergence warnings occur.
Tolerance (PLSRegression)
Convergence tolerance for iterative updates.
Lower tolerance can improve precision but may require more iterations.
Components (FastICA)
Number of independent components to estimate.
Use None to infer based on feature count.
Algorithm (FastICA)
Parallel or deflation update strategy.
parallel estimates components together; deflation extracts one by one.
Whiten mode (FastICA)
Whitening behavior before ICA optimization.
unit-variance is a robust default for regression pipelines.
Contrast function (FastICA)
Nonlinearity used to estimate non-Gaussian components.
logcosh is usually stable; try others when decomposition quality is poor.
Max iterations (FastICA)
Maximum iterations before stopping.
Increase if ICA fails to converge.
Tolerance (FastICA)
Stopping tolerance for ICA updates.
Lower values enforce stricter convergence.
Random state (FastICA)
Seed for ICA initialization.
Fix this value for reproducible components.
Components (FactorAnalysis)
Number of latent factors.
Use None to infer from input dimensionality.
SVD method (FactorAnalysis)
Backend used during factor estimation.
randomized is faster on large data; lapack is deterministic.
Iterated power (FactorAnalysis)
Power iterations used in randomized SVD.
Higher values can improve approximation quality.
Rotation (FactorAnalysis)
Optional factor rotation for interpretability.
Use None for unrotated factors, or varimax/quartimax for rotated solutions.
Tolerance (FactorAnalysis)
Convergence tolerance.
Lower values require stricter convergence.
Random state (FactorAnalysis)
Seed for randomized solver components.
Set for reproducible randomized runs.
Components (TruncatedSVD)
Number of singular vectors to keep.
Higher values preserve more information but reduce compression.
Algorithm (TruncatedSVD)
Solver for truncated decomposition.
randomized is efficient for large matrices; arpack can be more precise.
Power iterations (TruncatedSVD)
Additional iterations for randomized solver accuracy.
Increase when singular value gaps are small.
Tolerance (TruncatedSVD)
Convergence tolerance for ARPACK solver.
Mostly relevant when using arpack.
Random state (TruncatedSVD)
Seed for randomized solver.
Fix to make randomized decompositions reproducible.
Inputs
Target feature
Choose the numeric column you want the models to predict.
Select exactly one target feature. The algorithms will try to predict this value from the input features.
- Only one target can be active at a time; use the dropdown to switch quickly.
- Pick a continuous column (price, temperature, throughput) for best results.
- Changing the target automatically updates default stratification choices where possible.
Feature selection
Optionally enable automatic selectors before training.
Feature selectors reduce the input columns to the most informative subset.
- Choose one or more selectors to compare, or leave all unchecked to train with every selected feature.
- Each selector may expose hyperparameters in the Hyperparameters section below.
- Combining selectors and models multiplies the number of experiment runs.
Dimensionality reduction
Optionally project features into a lower-dimensional representation before modeling.
Dimensionality-reduction methods transform your selected input features before the regression model is trained.
- You can select multiple methods; each selected method multiplies the total number of runs.
- Methods run after feature selection and before the regression model.
- Use this when you have many correlated or noisy inputs, or when you want a compact representation.
PCA
Linear projection that keeps directions with the highest variance.
PCA transforms correlated features into orthogonal components that preserve as much variance as possible.
- Good baseline reducer for high-dimensional, correlated input sets.
- Improves numerical stability for some regressors.
- Component axes are not directly interpretable as original features.
PLSRegression
Supervised projection that uses both inputs and target to build latent variables.
PLSRegression finds latent components that maximize covariance between features and the target.
- Useful when predictors are highly collinear and target signal is weak.
- Unlike PCA, it is supervised and can focus on prediction-relevant structure.
FastICA
Independent component analysis for separating non-Gaussian latent sources.
FastICA estimates statistically independent components instead of variance-maximizing ones.
- Can help when signal sources are mixed and non-Gaussian.
- More sensitive to scaling and solver settings than PCA.
FactorAnalysis
Latent-factor model that explains observed variables through shared factors plus noise.
FactorAnalysis models each feature as a combination of latent factors and feature-specific noise.
- Useful for denoising and identifying compact latent structure.
- Common in domains where measurement noise is significant.
TruncatedSVD
Low-rank SVD projection that works well with sparse or large matrices.
TruncatedSVD projects features onto a smaller number of singular vectors without centering.
- Efficient for large feature spaces and sparse data.
- Often used as a scalable alternative to PCA in high dimensions.
Regression models
Pick the algorithms to evaluate for this experiment.
Select one or more models to train on the same dataset and compare their results.
- Toggle multiple options to run them in parallel; results appear in the runs table, predictions table, and charts.
- Model-specific settings are available under Hyperparameters → Models.
- For quick baselines keep a simple model (e.g., Linear Regression) alongside more complex methods.
Test Split
Hold-out test set
Reserve part of the data for final evaluation.
Enable the test set to keep a portion of data untouched during training and cross-validation.
- Disable when you want to use every row for cross-validation only.
- Keep it enabled to report unbiased metrics on unseen data.
- The reserved rows are excluded from model fitting.
Test size
Choose the fraction or fixed number of records set aside for testing.
The test size sets how much of the dataset is held out.
- Use fractions like 0.2 to reserve 20% of rows.
- Use integers like 200 to reserve a fixed number of rows.
- Typical fractional values range from 0.2 to 0.3 for balanced datasets.
- Use a smaller fraction when data is scarce, or larger when you expect heavy model tuning.
- Values outside the slider range can be typed directly when necessary.
Test split strategy
Control how rows are separated into train and test sets.
Pick a strategy that matches the nature of your data.
- Random: shuffle rows before splitting into train/test according to the chosen size.
- Time ordered: keep chronological order by placing the earliest rows in training and later rows in testing.
- Stratified: preserve the distribution of the selected stratify feature in both sets.
Strategies that rely on stratification will use the selected feature and binning settings below.
Test stratification
Balance the test split using a target or helper feature.
Select a feature to stratify the test split when using the Stratified strategy.
- Use the target for imbalanced regression targets or choose another categorical field.
- Continuous targets are discretised into bins using the setting below before splitting.
- Leave empty to disable stratification even if the strategy allows it.
Stratify bins
Set how many buckets to create when stratifying continuous values.
When the stratify field contains continuous numbers, the values are bucketed into bins to enable stratified sampling.
- Higher bin counts capture more detail but require larger datasets to keep buckets populated.
- Start with 5 bins for most targets; increase if you have plenty of records.
- Has no effect when stratifying by categorical features.
Forecasting
Forecasting
Create time-series forecasts from historical data (feature currently disabled in default builds).
The Forecasting tab focuses on time-aware modeling.
- Configure horizon and rolling-window validation settings for temporal data.
- Train forecasting models powered by scikit-learn and compare error metrics.
- Visualize predicted vs. actual values to validate performance.
This feature exists in the codebase but is currently disabled in both development and release tab flags by default.
When enabled, use it for scenarios like demand planning, capacity prediction, or financial projections.
Time Series Forecasting
Predict future values based on historical time series data.
Forecasting analyzes time-ordered data to predict future values.
Key concepts in time series forecasting:
- Trend - Long-term increase or decrease in the data
- Seasonality - Regular patterns that repeat over fixed periods
- Lag Features - Past values used to predict future ones
- Rolling Statistics - Moving averages and other window-based features
FastData uses forecasting models with time-series feature engineering and split strategies for experiments when the tab is enabled.
The previous sktime version is kept as a non-active reference in backend/services/legacy_forecasting/forecasting_service_sktime.py.
Common use cases: sales forecasting, demand prediction, financial projections.
Controls
Forecast horizon
How many future time steps to predict in each run.
The horizon sets how far ahead each model predicts.
- Use a small horizon (e.g., 1–24 steps) for near-term monitoring.
- Larger horizons extend the forecast further but usually increase uncertainty.
- Match the value to your reporting cadence (steps follow the dataset's sampling).
Models will generate this many points for every feature you selected.
Window strategy
Pick how training and validation windows move through the series.
Choose the cross-validation style used while fitting time-series models.
- Single Train/Test Split – one hold-out split; fastest baseline.
- Sliding Window – fixed-size training window that slides forward to mimic rolling forecasts.
- Expanding Window – training window grows over time to reuse all past observations.
Sliding windows are good when concept drift is likely; expanding windows favor stability.
Initial window size
Length of the first training window (auto when set to Auto).
Controls how many observations are used before the first forecast window.
- Auto picks a reasonable size based on the horizon and model defaults.
- Provide a positive integer to force a specific training length.
- Use smaller windows for rapidly changing signals; larger windows for long-term patterns.
This applies to sliding and expanding strategies; the single split uses the full training span.
Hyperparameters
Strategy (naive)
How the naive forecaster projects future values.
Select the baseline rule used to forecast.
- last: repeats the last observed value.
- mean: forecasts the historical mean.
- drift: extends the overall linear trend.
Seasonal period (naive)
Length of the seasonality cycle used by the naive model.
Defines how many time steps make up one season.
- Use 7 for weekly seasonality on daily data, 12 for monthly seasonality on monthly data, etc.
Seasonal period (theta)
Season length used by the theta forecaster.
Helps the model separate seasonality from trend when present.
- Set to the number of observations per season.
Deseasonalize (theta)
Remove seasonality before fitting the theta model.
Enable when your series has strong seasonal patterns.
- If disabled, the model fits directly on the original data.
Trend (exponential smoothing)
Type of trend component to include.
Controls how the model extrapolates the long-term movement.
- add: linear trend (adds the trend each step).
- mul: multiplicative trend (scales with level).
- None: no explicit trend component.
Seasonal (exponential smoothing)
Seasonality mode for the exponential smoothing model.
Select how seasonal patterns are modeled.
- add: seasonality adds to the level.
- mul: seasonality multiplies the level.
- None: no seasonal component.
Seasonal period (exponential smoothing)
Number of steps in one seasonal cycle.
Use the length of your repeating pattern to get better seasonal fits.
Polynomial degree
Complexity of the polynomial trend.
Higher degrees allow more curvature in the trend.
- Start with a low degree to avoid overfitting.
Use Box-Cox (BATS)
Apply a Box-Cox transform to stabilize variance.
Helps when variability grows with the level of the series.
Use trend (BATS)
Include a trend component in the BATS model.
Enable to capture long-term upward or downward movement.
Use damped trend (BATS)
Dampen the trend so it flattens over time.
Useful when trends should level off rather than grow indefinitely.
Use Box-Cox (TBATS)
Apply a Box-Cox transform to stabilize variance.
Improves modeling when variance changes with the level.
Use trend (TBATS)
Include a trend component in the TBATS model.
Enable to track long-term movement in the series.
Use damped trend (TBATS)
Dampen the trend component over time.
Prevents overly aggressive trend extrapolation.
Smoothing (Croston)
Smoothing factor for intermittent demand forecasting.
Controls how quickly the model reacts to new observations.
- Lower values smooth more, higher values respond faster.
Alpha (time series ridge)
Regularization strength for ridge regression in time series.
Higher alpha shrinks coefficients and reduces overfitting.
Window length (time series ridge)
Number of past steps used as features.
Longer windows capture more history but increase model size.
Strategy (time series ridge)
How multi-step forecasts are generated.
Choose between recursive, direct, or multioutput strategies.
- recursive: one-step model iterated forward.
- direct: separate model per horizon step.
- multioutput: single model predicts all steps.
Alpha (time series lasso)
Regularization strength for lasso regression.
Higher alpha encourages sparse coefficients.
Window length (time series lasso)
Number of lagged steps used as input features.
Increase to capture longer temporal dependencies.
Strategy (time series lasso)
Approach for multi-step forecasting.
Recursive, direct, or multioutput behavior for generating horizons.
Estimators (time series random forest)
Number of trees in the forest.
More trees improve stability but increase runtime.
Max depth (time series random forest)
Maximum depth of each tree.
Use None to allow trees to expand fully.
Window length (time series random forest)
Number of historical steps turned into features.
Larger windows capture more history but can add noise.
Strategy (time series random forest)
How the model produces multi-step forecasts.
Select recursive, direct, or multioutput forecasting.
Random state (time series random forest)
Seed for the random forest training process.
Set a fixed seed to make runs reproducible.
Estimators (time series gradient boosting)
Number of boosting stages.
More stages can improve accuracy but increase risk of overfitting.
Max depth (time series gradient boosting)
Depth of individual regression trees.
Shallower trees generalize better on noisy data.
Learning rate (time series gradient boosting)
Shrinkage applied to each boosting step.
Smaller values require more estimators but often generalize better.
Window length (time series gradient boosting)
Number of lagged steps used as input features.
Increase to capture longer temporal patterns.
Strategy (time series gradient boosting)
Multi-step forecasting method for gradient boosting.
Choose between recursive, direct, or multioutput forecasting.
Random state (time series gradient boosting)
Seed for model training randomness.
Set to ensure reproducible results across runs.
Layers (time series MLP)
Hidden layer sizes for the MLP regressor.
Enter comma-separated sizes, e.g. 64,32,16.
More layers increase capacity but also training time.
Activation (time series MLP)
Activation function in hidden layers.
relu is a good default; tanh can smooth outputs.
Solver (time series MLP)
Optimizer for MLP training.
adam works well for most datasets.
lbfgs can be faster on smaller datasets.
Alpha (time series MLP)
L2 regularization strength.
Higher values apply more weight decay to reduce overfitting.
Learning rate (time series MLP)
Learning rate schedule.
constant keeps a fixed rate.
adaptive reduces the rate when learning stalls.
Max iterations (time series MLP)
Maximum training epochs.
Increase if the model fails to converge.
Window length (time series MLP)
Number of lagged steps used as features.
Longer windows capture more history but increase model size.
Strategy (time series MLP)
How the MLP produces multi-step forecasts.
Recursive, direct, or multioutput forecasting.
Random state (time series MLP)
Seed for weight initialization.
Set to ensure reproducible results.
Models
Forecasting models
Pick one or more algorithms to run for each selected feature.
Select the models to include in the experiment. Each option offers different strengths:
- Naive – simple baselines using the last value, mean, or drift trend.
- Theta – decomposes the series and extrapolates a flexible trend component.
- Exponential Smoothing – configurable trend/seasonality smoothing (additive or multiplicative).
- Polynomial Trend – fits a low-degree polynomial to capture smooth long-term movements.
- BATS – Box-Cox transform with ARMA errors, trend, and seasonality support.
- TBATS – BATS extended with multiple seasonal periods and trigonometric components.
- Croston – designed for intermittent demand with sporadic non-zero observations.
- Time Series Ridge/Lasso Regression – reduction to tabular regression with sliding windows.
- Time Series Random Forest – ensemble of decision trees for non-linear relationships.
- Time Series Gradient Boosting – boosted trees for strong accuracy on complex patterns.
- Time Series MLP – neural network regressor for nonlinear temporal patterns.
Run several models together to compare metrics and pick the best-performing approach.
Target feature
Optional exogenous target for time series regression models.
Select a target feature when using regression-based forecasters (ridge, lasso, random forest, gradient boosting).
- Leave blank for univariate forecasts where each feature predicts itself.
- Choose a single target to train models that learn relationships from multiple inputs to one output.
- The target is included in preprocessing so scaling or filtering stays consistent.
Only one target can be active at a time; deselect it to return to univariate mode.
Chat
Chat sessions
Create, switch, and delete independent chat histories.
Chat history is stored per session.
- New Chat creates a separate empty session.
- Delete Chat removes only the currently selected session.
- Other sessions and their history remain available in the session selector.
Using Chat
Ask the built-in assistant and keep a record of conversations.
The chat window lets you send questions to the assistant and keeps a history in the log database.
- Chat messages are saved automatically with the log database.
- Use the stop button next to Send to cancel an in-progress streamed reply.
- Each chat session can be deleted separately without removing other log entries.
- Resetting the log database clears chat history too.
Settings
LLM provider
Choose the backend that powers chat replies.
Select which provider handles the chat requests sent from the chat window.
- ChatGPT (OpenAI) uses OpenAI-hosted models and requires an API key.
- Ollama (Local) runs models on your machine without an API key.
Helpful links:
- ChatGPT and OpenAI API
- Ollama and install instructions
If you want more guidance, use Ask from AI below to send this topic to the chat window.
Model name
Enter the model identifier for the selected provider.
Provide the exact model name your provider expects.
- OpenAI examples:
gpt-4o-mini,gpt-4o,gpt-3.5-turbo - Ollama examples:
llama3.2,mistral,codellama
Browse model lists:
Use the refresh icon next to Model name to fetch available models for the selected provider.
You can click Ask from AI below for help choosing a model.
Thinking mode
Control how much intermediate reasoning the model exposes in chat.
Choose how much model reasoning is shown in the assistant reply card.
- Off: hides reasoning output and focuses on final answers.
- Standard: balanced reasoning visibility for normal use.
- High: requests more extensive reasoning when the provider supports it.
Provider behavior varies by model. If a model ignores this setting, chat still works and returns normal replies.
OpenAI API key
Add your OpenAI key for ChatGPT/OpenAI access.
This key is only required for the ChatGPT (OpenAI) provider.
Create or manage keys at OpenAI API keys.
If you do not have a key yet, sign in and generate one, then paste it here.
Use Test to save the key and verify it by fetching the available OpenAI models.
Use Ask from AI below if you want step-by-step guidance.