Introduction
...
Key Features
Approach
The Forecast Accelerator follows a structured approach
...
. It starts by mapping the data and incorporating as many features as possible to provide comprehensive context for the forecasting model. This includes additional sources that might influence future values, such as special events.
...
A preliminary forecasting model is built and tested over a holding period to determine the best model parameters and assess accuracy.
...
The model is trained with the latest available data to capture the most recent trends.
...
It predicts revenue or quantity for the upcoming periods.
...
The forecasts are exported in a data source format similar to the input data, making them easy to integrate into other processes.
Outputs
The primary outputs of the Forecast Accelerator are:
Revenue or Quantity Forecasts: These can be daily, weekly, or monthly.
Charts: Various charts are available to visualize the outputs, with options to filter at different levels of granularity.
Limitations
The document also highlights several limitations:
Best Practices
To ensure the Forecast Accelerator functions optimally, it's important to follow certain best practices. These include using complete and relevant features, ensuring the training period is fully covered, and pre-processing data for new or rarely sold products.
Attributes/Features: Ensure that all relevant features are included and complete. The accuracy of the forecasts depends on having the right features
...
, so it's crucial to avoid incomplete or irrelevant data.
Training Periods:
...
Use a complete training period
...
. Partial data, such as half a week, can bias the results, so make sure the training period is fully covered.
Sparse Data:
...
For new or rarely sold products,
...
aggregations and pre-processing can help
...
improve forecast accuracy. Ensure that past sales data is as comprehensive as possible.
Limitations
While the Forecast Accelerator is a powerful tool, there are some limitations to be aware of. Incomplete or irrelevant features, partial training periods, and sparse data can affect forecast accuracy. Additionally, the model does not account for long-term contracts or agreements
...
and requires custom code for specific features, which can complicate updates.
...
Data Requirements: Specific data requirements must be met for optimal performance.
Creating a New Model
This guide will walk you through the steps to create a new model, configure it, train it, and export the forecast results.
Navigate to Models: Go to the Optimization > Models menu.
Add a New Model: Click to add a new model and select "Forecast" as the Model Class.
Initial Setup: When the model opens for the first time, it will run a small calculation to generate a parameters table.
Model Definition
Definition Step: Define the input data and their mapping. This includes selecting the data source and applying filters.
Definition Tab: Map the historical data and define the scope of the forecast.
Model Configuration Tab: Configure the model's settings, such as test set size, past time steps for lags, and rolling statistical windows.
Additional Configuration
Additional Sources Tab: Define up to three additional sources to extend the data or provide future values.
Aggregation Level Tab: Set the aggregation for training and the granularity for the final forecast.
Model Training
Training Process: The model will process the data into a time series, enrich it with extra features, and split it into training and test sets.
Results Tabs:
Model Training Results: Summarize the model's performance.
Train and Test Forecasts: Display forecasts alongside historical metrics.
Training Curves: Show diagnostic metrics.
Elasticity Settings: Configure settings for elasticity calculation if applicable.
Model Predictions
Overview Tab: Display overall results for all products.
Details Tab: Show detailed forecasts at the defined granularity level.
Export Forecast
Export Parameters: Define the parameters for exporting the forecast data to a Data Source.
Final Export: Save the forecast data at the product, customer, and date level.
Additional Information for Business Users
Key Recommendations
Automatic Filtering: Keep the automatic filtering of negative quantities and prices enabled for more accurate predictions.
Log Transformation: Consider using log transformation for metrics with skewed sales data.
Model Tuning: Perform automatic model tuning to improve results, especially for large datasets.
Aggregation Levels: Choose appropriate aggregation levels to balance training speed and result precision.
Best Practices
Data Consistency: Ensure data consistency between the Definition step source and additional sources to avoid errors during training.
Feature Selection: Include relevant categorical and numerical features to enhance the forecast accuracy.
Elasticity Calculation: Use elasticity settings to understand the impact of price changes on quantity forecasts.
Troubleshooting
Model Performance: If the model performance is poor, consider increasing the number of tuning trials or adjusting the aggregation levels.
Export Issues: Ensure the Data Source keys are correctly defined to avoid inconsistencies during export.
By following this guide, business users can effectively utilize the Forecast Accelerator to generate accurate and actionable forecasts, aiding in better decision-making and strategic planning.
The table lists the data columns required and optional for running a forecast model. Each column represents a type of data that the model needs to make accurate predictions.
How to Use This Information
Collecting Data:
Ensure you have at least 2 years of transaction history.
Gather data for each required column listed in the table.
Understanding Each Column:
Date: This is the date of each transaction. It should be defined as a dimension in your data source or datamart.
Product: This is typically the Product ID. It identifies the product involved in each transaction.
Revenue: This is the total revenue from each transaction, calculated as the end-customer price multiplied by the quantity sold. Avoid negative or null values.
Quantity: This is the number of units sold in each transaction. Avoid negative or null values. Using the log of the quantity can help if you have a wide range of quantities.
Store: Although required, this column is not used in the current version and will be removed in the next version.
Optional Columns:
Revenue at List Price: This is the list price multiplied by the quantity. It helps calculate discount rates and is recommended if available.
Product Categorical Features: These are categorical attributes related to the product that might influence sales, such as Product Category, Competitor Name, Product Life Cycle, and Promotion Tags.
Product Numerical Features: These are numerical attributes that can be averaged over time, like product ratings or average sales price.
Customer: This is required only if you choose to add a customer dimension. It includes customer-related data.
Customer Categorical Features: These are categorical attributes related to the customer that might influence sales, like Customer Segment or Loyalty Status.
Setting Data Filters:
Ensure your data filter is set to capture complete periods (e.g., full weeks or months) to avoid partial data that could skew the forecast.
Practical Steps
...
Data Preparation:
Gather and clean your data according to the requirements listed.
Ensure all required columns are included and properly formatted.
...
Model Configuration:
Input the prepared data into your forecast model.
Configure the model settings, including any optional columns that might improve the forecast accuracy.
Running the Model:
...
Execute the forecast model with the prepared data.
...
Understanding these limitations will help you manage expectations and make the most of the tool.
Attributes/Features: Incomplete or irrelevant features can lead to poor output quality.
Training Periods: Partial data can bias the results.
Sparse Data: Forecasts might be less accurate for new or rarely sold products.
Long-term Contracts/Agreements: The model does not use long-term contracts or agreements stored in the solution.
No Predefined Extension Points: Custom code is required to add specific features, which can complicate updates.
Data Requirements: Specific data requirements must be met for optimal performance.
| Info |
|---|
Check out the Data Requirements for this accelerator /wiki/spaces/ACCDEV/pages/5768314910 |