Sales forecasting with spreadsheet integration
How deep learning can solve real business problems directly from your spreadsheets
This tutorial will show you how to use the full power of the Peltarion Platform in a real world situation.
You will build a model that predicts daily sales revenue from many parameters. The platform lets you deploy this model for production, allowing you to directly integrate predictions in your Google Sheets or integrate predictions in your Microsoft Excel spreadsheets.
- Target audience: Beginners
- Estimated time: Setup - 10 min | Training - 5 min | Deployment - 1 min
You will learn to
- Build a deep learning model with no code.
- Predict sales numbers from spreadsheet data.
- Deploy your model for production on the Peltarion Platform.
- Integrate your model with Google Sheets or Microsoft Excel.
Create a project
Let’s begin! Log in to the Peltarion Platform and click New project. Name the project, e.g. Sales, so you know what kind of project it is.
Import the data
After creating the project, you will be taken to the Datasets view, where you can import data.
Use our data library
Click the Data library button and look for the Sales forecasting - tutorial data dataset in the list. Click on it to get more information.
If you agree with the license, click Accept and import. This will import the dataset in your project, and you will be taken to the dataset’s details where you can edit features and subsets.
There are few companies that openly share revenue data. So for this tutorial, we’ve generated a synthetic dataset that looks like the real thing but is free to use.
The dataset contains daily values for many different shops over a couple of years, such as the date, national holidays, type of shop, ongoing advertisement and promotion campaigns.
The shop’s daily revenue is also given, which is what the model will learn to predict.
The appropriate feature encoding is selected automatically:
Click Save version. The data is now ready to be used in an experiment.
Build the model
To begin an experiment click Use in new experiment. The Experiment wizard dialog will open.
The platform creates subsets automatically from the imported data.
The Training subset is used by the model to improve its predictions.
The Validation subset isn’t shown to the model while training. You can use it to evaluate how well the model performs on data it has never seen before.
Click Next to set up the model’s inputs and target.
Input(s) / target tab
In the Inputs column, select everything EXCEPT the Date, Year, and Revenue (since it’s the target).
We won’t use specific time information, like Date and Year, to train the model because we want to make predictions for any future (or long past) date.
Features like the week number provide enough information about the time period, i.e., if an example is from winter or summer, while being general enough to work for any future year.
Select Revenue in the Target column. The target is what the model will learn to predict.
Given the inputs and target selected, the wizard automatically recommends Tabular regression as Problem type, and selects Tabular as the Recommended snippet.
Click on Create, and the wizard will create a model that fits your tabular data.
Everything is set in the Modeling view thanks to the wizard. Click Run to start training your model.
As the model trains, you can follow its performance in the Evaluation view.
Evaluate your model
The Evaluation view shows you how the model performance improves as the training progresses. The loss and metrics plot gives an overall idea of the training process. You can also check the scatter plot in the Predictions inspection tab to see predictions of specific examples.
Training will stop automatically when the model stops improving thanks to early stopping. The experiment status will change from Running to Early stopped, and you can move on to the next section to deploy your model.
In this tutorial we’ll build a single model.
However, you can easily Duplicate your experiment to try different model configurations. The tutorial How to improve a model that uses tabular data goes into more details about how to proceed.
Deploy your model
Navigate to the Deployment view of the platform, and click on New deployment.
Select the Experiment that you want to deploy.
An experiment corresponds to a particular model trained with specific settings. There should only be one available, unless you have experimented with different models in your project.
Select the Checkpoint you want the deployed model to use.
Checkpoints are made throughout training. Select the Best checkpoint, since this is when the model had the best performance.
Click Enable to let your deployment go live! You can now send requests to your model from anywhere, anytime.
Integrate your model where you work
When a model is deployed, the Peltarion Platform makes it available to you from anywhere with an HTTP connection. This means that you can use the Deployment API to get predictions in your applications.
You can use the model directly inside your spreadsheets. To learn how, check out one of these short tutorials:
Congratulations, you have now completed the sales forecasting tutorial!
You’ve built a simple model which analyzes a large amount of tabulated data to solve a regression problem.
This dataset is very simple, but much more can be done with the use of deep learning. Deep learning for sales forecasting could for example use combinations of complex data types like images and text. To see how to solve problems using other types of data, check out our other tutorials.