Forecast Accelerator deploys a powerful machine learning approach to predict future sales, either forecasting revenue or quantity for the next periods.
Approach
Forecast Accelerator relies on several steps:
-
First, map the data and include as many features as possible to give enough context and information to the forecasting model, including additional sources that could provide future values, such as special events.
-
Build a first forecasting model and test outputs for a holding time frame in order to define the best model parameters and get metrics on the holding period to get a fair assessment of the model accuracy.
-
Train the model with the latest available data in order to get full knowledge of latest trends.
-
Predict revenue and quantity for next periods.
-
Export forecast in a Data Source (optional), with similar structure with input data so it can easily used in other processes.
Outputs
The results of a Forecast model are revenue or quantity for the next periods (daily, weekly or monthly). Last price is used to compute quantity (if revenue is used as the metric) or revenue (if quantity is used as the metric). Several charts are also available to check outputs,with possibility to filter at different level of granularity.
Limitations
-
Attributes / features – Multi-factor elasticity relies on having the right features that impact the sold quantity. If those features are not available or the data are incomplete, output quality will be low. Also, additional features should either be product or data related, otherwise each additional feature would create a specific time series.
-
Training periods – Training period should be a complete period, meaning if the last period only contains partial data, like half a week, the outputs will be biased and probably wrong.
-
Sparse data – Forecasting relies on past revenue/sold quantity from past transactions to forecast coming sales. However, if data is sparse, meaning there is low recurrence of transactions for the same product/customer, forecasts might be less accurate. To work around this potential issue, please use aggregations built within the model that could help a lot to improve outputs. Combinations of product/customer that are new or barely happened in the past will also be less accurate.
-
No use of the long term contracts/agreements – This forecast model leverage machine learning based on the past transactions and patterns that can be found in those transactions, but not using the contracts or agreements stored in the solution.
-
No predefined extension point – There is no out-of-the-box extension point defined for now. If you intend to add specific features, custom code should be written. (But then the accelerator becomes specific and it cannot be updated without extra effort to port those modifications.) Do not hesitate to report specific requirements and possible extension points to Pricefx.
-
Data requirements – See Data Requirements (Optimization - Forecast)