Experiments
An experiment represents a specific forecasting task in which the user defines the dataset(s) to use, selects the target variable, and specifies the forecast horizon. Once the parameters are configured and the experiment is run, the results are generated and made available for review. The results page provides insights into how the forecasts were produced, evaluates performance, and allows users to analyze their findings in depth.
The Experiments page is designed to feel familiar to users of the Datasets page, ensuring continuity in the Studio experience, while also clearly highlighting the unique purpose and workflow of Experiments compared to Datasets.
Experiments Table Overview
The Experiments table provides a centralized view of all your experiments and their current statuses. Key features include grouping, sorting, and status indicators.
Grouped vs. Ungrouped
By default, the table is in Grouped mode, allowing you to clearly separate experiments into categories. You can also toggle to Ungrouped view to see all experiments in a single list, regardless of their group.
Sorting Experiments
The Sort By dropdown lets you organize the table according to specific criteria, such as the status of an experiment, creation date, or other available attributes.
Experiment Statuses
Studio uses badges to indicate the current state of each experiment. The possible statuses are:
Completed - The experiment has successfully finished. The model was built without errors, and all results are now available for review.
Failed - An error occurred during the experiment process. The model was not built, and results are unavailable.
In Progress - The experiment is currently running. Model training and evaluation are actively underway.
Queued - The experiment request has been submitted and is waiting in the queue. It will begin automatically once resources are available.
Creating an Experiment
To create a new experiment, from the experiments page click the Create Experiment button. This will launch the Create Experiment popup. Creating an exerpeiment is splitg into three key steps with a forth step to validate and summarise before creation.
- Selecting a Dataset
- Selecting Target and Predictors
- Configuring the Model Settings
- Summary of Experiment
Selecting a Dataset
In the first step of creating an experiment, you must select the dataset or datasets that you wish to use for this experiment. Any existing datasets are shown in table, but you can add a new dataset through the Upload new dataset popup. For each dataset, its name, number of rows and columns, start and end date are listed so that you can visually select similar datasets. As well as this, you can generate a quick visualisation of each dataset via the Plot button.
Once you have selected the datasets for the experiment, the statistics panel provides a comprehensive overview of the experiment dataset. It displays the sampling period, along with the first and last timestamps. You can render a full plot of the experiment dataset to visualize your selected data. Below the plot button, a table presents key per-column statistics, including the minimum, maximum, mean, and standard deviation for each column.
Selecting Target and Predictors
The second stage of creating an experiment involves the selection of your target and predictor columns. The target is the column which you want to forecast for, for example temperature and load. The remaining columns then become the predictors to which are used to forecast for the target. Within this stage, Studio allows you select the target, toggle the predictors and then also see the data horizons for each predictor in relation to the target.
Alongside selecting the target and predictors, within this section you also specify the forecast horizon and the Training and Validation ranges. The forecast horizon is how far in the future you want to generate predictions for, for example, 1 minute or 7 days. Studio supports five types of period:
- Sample - Forecast by number of data points (e.g., predict the next 10 samples)
- Minute - Forecast by minutes into the future (e.g., predict the next 30 minutes)
- Hour - Forecast by hours into the future (e.g., predict the next 24 hours)
- Day - Forecast by days into the future (e.g., predict the next 7 days)
- Week - Forecast by weeks into the future (e.g., predict the next 4 weeks)
In time-series forecasting, the dataset must be split chronologically to preserve the natural flow of time. The training set contains an older portion of the sequence, which the model uses to learn temporal patterns and trends. The validation set is composed of more recent data so the model can be evaluated on how well it predicts the future rather than the past. This prevents data leakage, ensures realistic evaluation, and reflects the real-world scenario where future values are unknown during training.
Studio offers three ways to define your training and validation data. First, you can manually set the exact start and end dates for both the Training and Validation ranges, which allows for precise and informed control over the data. Second, you can select from multiple predefined ratios for the Training and Validation data using the Ratio buttons, providing a fast and easy-to-understand method for splitting your data. Third, you can define the Training and Validation ranges graphically using a visual representation of your data. Clicking on Select on Graph opens a pop-up window where your time-series data is displayed. You can then click directly on the graph to set the start and end points for both ranges. Below the time-series graph, the Range Slider allows you to zoom into the data by adjusting the date range manually, providing a simple and hands-on way to visualize and interact with your data.
Defining the Model Settings
After selecting your target and predictors, the next step is to define and configure your model. Every experiment must have a name, and you can optionally assign it to a group and provide a description. Grouping your experiments helps organize and categorize your work, making it easier to manage multiple experiments. Experiments without a group will appear as Ungrouped in the Grouped view of the Experiments page. Adding a description allows you and your team members to quickly distinguish between different experiments, providing additional context and clarity.
At the core of an experiment lies the mathematical settings that determine how Studio will use your data to generate forecasts. For simplicity, Studio automatically selects the most appropriate settings for your experiment, allowing you to get started quickly without needing to configure every option manually. These settings are grouped into three categories:
-
Basic - Includes fundamental options such as Target and Predictor Offsets, Data Normalisation, and Daily Cycle. These are automatically preconfigured, but can be adjusted them to meet specific requirements.
-
Advanced - This section includes options that control the deeper behavior of the model, such as the maximum feature depth, how aggressively offsets and past data points are used, and the confidence interval for predictions. You can adjust these settings to influence model complexity, control how historical data impacts forecasts, and set the level of certainty for prediction intervals.
-
Transformations - These define how raw data is transformed into features for the model, such as moving averages, Fourier components, or categorical encodings. Studio can automatically select the most relevant transformations, but you can manually enable or disable individual transformations if you want to incorporate domain knowledge or experiment with specific feature types.
In short, most users can rely on the automatic defaults for robust forecasting, while the ability to configure individual settings provides flexibility for advanced use cases or custom modeling requirements. This balance ensures that experiments are both easy to run and fully customizable when needed. Below is an explanation of each of the mathemtical settings that Studio provides:
Basic Mathematical Settings
Basic Mathematical Settings define how the forecasting model interprets past data, selects relevant inputs, and structures the prediction task. It controls the use of historical information, the specified prediction horizon, and the handling of recurring patterns, allowing the model to capture temporal dependencies. The Basic Mathematical Setttings form the foundation of the model’s understanding of the time-series, ensuring that forecasts are accurate, stable, and interpretable.
Advanced Mathematical Settings
Advanced mathematical settings provide additional control over model behavior, uncertainty estimation, and feature selection. While the basic settings define how the model interprets and structures the time-series, advanced settings let users fine-tune the forecasting process, manage risk, and optimize model complexity for specific applications. Configuring these options enables greater precision, robustness, and interpretability in more complex or high-stakes scenarios.
Transformations
Transformations are systematic ways of modifying or encoding time-series data to make patterns more visible, regular, or easier for models to learn. They help to uncover underlying structure by emphasizing trends, seasonality, cycles, or other predictable patterns, while reducing noise or irregularities.
- Exponential Moving Average
- Simple Moving Average
- Trend
- Polynomial
- Piecewise Linear
- Day of Week
- Rest of Week
- Month
- Fourier
- Perioidic
- Intercept
- Identity
Summary of Experiment
The final stage of this process provides a concise overview of your experiment, allowing you to review and verify each step in a simple, intuitive manner. Once you have confirmed that everything is configured correctly, click the Create button to run the experiment and view the results.
Iterating on an Experiment
Iterating on an experiment is a process similar to updating a dataset on the Datasets page. After reviewing the results of a completed experiment, you may want to make adjustments, such as modifying individual parameters, changing the training or validation ranges, or including or excluding certain data columns. Iteration allows you to refine your model and explore how different configurations affect results.
The process for iterating an experiment follows the same steps as creating the original experiment. However, instead of overwriting the existing experiment, the new iteration is added as a separate experiment in the Experiments table. This approach allows you to track changes over time, compare multiple iterations, and maintain continuity across experiments while preserving previous results for reference. By iterating in this way, you can test different ideas and assumptions whilst ensuring that any and all changes are documented and reproducible.