Skip to content

Latest commit

 

History

History
288 lines (196 loc) · 11.3 KB

PredictIncomingCashflow.md

File metadata and controls

288 lines (196 loc) · 11.3 KB
Raw

Predict Incoming Cashflow

In this section we'll create a model to predict incoming cashflow based on historical payment delays for previous sales.

We'll be using Azure Machine Learning for this.

Setup in Synapse

In Synapse Studio, we will create a view joining data coming from SalesOrderHeaders and Payments tables that will be used for the prediction. You can create this view either via Synapse Studio or via Azure Data Studio.

  • Choose the Develop tab, select SQL Scripts and click on Actions then New SQL Script

Note : Ensure to connect to your SQL Pool

  • In the newly created script tab, copy paste the following SQL Query that will execute a join between SalesOrderHeaders and Payments to create a new view.
CREATE VIEW [dbo].[SalesPaymentsFull]
	AS SELECT s.[SALESDOCUMENT]
    , s.[CUSTOMERNAME]
    , s.[CUSTOMERGROUP]
    , s.[BILLINGCOMPANYCODE]
    , s.[BILLINGDOCUMENTDATE]
    , p.[PaymentDate] as PAYMENTDATE
    , s.[CUSTOMERACCOUNTGROUP]
    , s.[CREDITCONTROLAREA]
    , s.[DISTRIBUTIONCHANNEL]
    , s.[ORGANIZATIONDIVISION]
    , s.[SALESDISTRICT]
    , s.[SALESGROUP]
    , s.[SALESOFFICE]
    , s.[SALESORGANIZATION]
    , s.[SDDOCUMENTCATEGORY]
    , s.[CITYNAME]
    , s.[POSTALCODE]
    , DATEDIFF(dayofyear, s.BILLINGDOCUMENTDATE, p.PaymentDate) as PAYMENTDELAYINDAYS
 FROM [dbo].[SalesOrderHeaders] as s
JOIN [dbo].[Payments] as p ON REPLACE(LTRIM(REPLACE(s.[SALESDOCUMENT], '0', ' ')), ' ', '0') = p.[SalesOrderNr]

After Refreshthe view will appear under Views when using Azure Data Studio.

You can now test the view by executing the SQL:

select * from SalesPaymentsFull

Azure Machine Learning

The Azure Machine Learning Workspace is automatically deployed with the Terraform script from the first steps. If you didn't use the Terraform option you will first have to create an Azure Machine Learning Workspace via these instructions.

For more info on Azure ML, pleae have a look at What is automated machine learning (AutoML).

Open the ML Studio

By default the name for the ML Workspace is sap-data-ml-ws. Go to this workspace via your resource group or using the search bar. You can now open the ML Studio via Launch studio from here or alternatively sign in via https://ml.azure.com.

DataStore Creation

First you have to point the ML studio to the location of your data, which is the Synapse SQL Pool. For this you have to create a DataStore.

Go to the Datastores view and choose New datastore. As name we use sap_data_ml_ds. Connect to the Synapse database.

  • Datastore name : sap_data_ml_ds
  • Use Data store type : Azure SQL Database
  • Use Account Selection method : Enter Manually
  • Use your Synapse Workspace name as Server Name
  • Use your Synapse SQL Pool as Database Name, in our case this is sapdatasynsql
  • Select your Subscription
  • Enter your resource group name, in our case this is microhack-sap-data-rg
  • Authentication Type : SQL Authentication
  • Enter UserId and Password

Automated ML

We'll be using Automated Machine Learning to predict when customers will pay for their Sales Orders

  • On the left menu, click on Automated ML

  • Select New Automated ML Run
  • Select Create Dataset > From datastore

A Guided Procedure will appear :

  • Basic info : Provide a Name for the Dataset. We use SalesPaymentsView. Next.

  • DataStore Selection : Select your datastore.

  • Use the following SQL query to get all the data from the view defined above.
select * from SalesPaymentsFull
  • Settings and Preview : To Ensure that your query is working fine you are able to visualize the data in the next window.

  • Schema : In order to get a model we have to do some cleaning of the data.

  1. Uncheck the date fields (BILLINGDOCUMENTDATE, PAYMENTDATE) (We will not use these in our the model.)
  1. Uncheck the fields that do not contain any data or which are not relevant for the forecast. Eg. SALESDOCUMENT, SALESGROUP, SALESOFFICE
  • Confirm details Create the dataset

Configure the Automated ML Run

  • Select the newly created Dataset and create a new experiment.

  1. Specify a name, we use sap-data-ml-experiment
  2. Select the Target Column : in our case we will use PAYMENTDELAYINDAYS to predict the forecast.

  1. Create a new compute that will be used to train your model. As name we use sap-data-ml-vm. In this example we use a maximum of 3 nodes to increase the processing power.

Note : Create a cluster of for example 3 nodes. Azure ML can then run multiple trainings in parallel. This will reduce the runtime of the Automated ML run.

  • We can now select the ML task type we want to use for this experiment, as we want to build prediction on a numeric value we will select the Regression task type.

  • Then we need to configure the Regression using Additional Configuration settings.

    1. Select Normalized root mean squared error as Primary metric.
    1. In order to reduce the runtime of our 'Automated ML Run', we'll deselect some algorithms : ElasticNet, GradientBoosting, KNN, LassoLars, SGD, RandomForest, ExtremeRandomTrees, LightGBM, FastLinearRegressor, OnlineGradientDescentRegressor

    Note : If you have time you can include these algorithms.

    1. Save, Next, skip the optional steps and click on Finish
    1. During the run you can follow-up on the tested models via the Models tab

Deploy the best model

In this step we will deploy the best model that has been trained by AutoML and test it.

Note : the best model is selected based on the error between the predicted Payment Offset and the Actual offset. The model with the least error is selected. For more info on this, see How automated ML works

  • When the training is over, you can see the Best model summary section filled with the best algorithm, click on it.

  • You can navigate into the different sections and visualize the information about this algorithm, then click on deploy.

  • Specify a name for your deployment, we used sap-data-ml-model, and select Azure Container Instance as compute type.

Note: Select Deploy to WebService

  • Validate and wait for the completion of the deployment. This can take a few minutes.

  • When completed, click on the link to the Deploy status of the deployed model.

  • In this page, you will have access to information on your endpoint. It provides code samples to consume it from Python or C# but also a page to directly test your model.

Note: for a sample Python program, have a look at testForecast.py

  • The Azure ML can also be called as a REST Interface. You could use this REST Interace in a custom Fiori App or ABAP code to execute a Payment Prediction when creating a Sales Order.

Test the Payment Delay/Offset Prediction

Select the Test tab and insert values coming from the SalesPaymentsFull view created at the beginning to replace the example_value value for the different fields and run the model.

Note : Experiment with CustomerGroup Z1 and Z2 and note the Payment Delay/Offset. You can also compare with the actual value in powerBI or in Synapse.

You can now proceed with the next step.

[Optional] Test the ML Endpoint via HTTP

You can also test the ML model via http. You can find the URL to use via the ML endpoint menu.

Select your ML Model

Note : you can have a look at the API definition using the Swagger UI

You can test the ML model via a HTTP post request on the Endpoint URI. Below you can find a sample request.

HTTP Header

Content-Type : application/json

HTTP Body

{"data": [ 
    { 
    "CUSTOMERNAME": "Westend Cycles",
    "CUSTOMERGROUP": "Z1",
    "BILLINGCOMPANYCODE": 1710,
    "CUSTOMERACCOUNTGROUP": "KUNA",
    "CREDITCONTROLAREA": "A000",
    "DISTRIBUTIONCHANNEL": 10,
    "ORGANIZATIONDIVISION": 0,
    "SALESDISTRICT": "US0003",
    "SALESORGANIZATION": 1710,
    "SDDOCUMENTCATEGORY": "C",
    "CITYNAME": "RALEIGH",
    "POSTALCODE": "27603"
    },
    { 
    "CUSTOMERNAME": "Skymart Corp",
    "CUSTOMERGROUP": "Z2",
    "BILLINGCOMPANYCODE": 1710,
    "CUSTOMERACCOUNTGROUP": "KUNA",
    "CREDITCONTROLAREA": "A000",
    "DISTRIBUTIONCHANNEL": 10,
    "ORGANIZATIONDIVISION": 0,
    "SALESDISTRICT": "US0004",
    "SALESORGANIZATION": 1710,
    "SDDOCUMENTCATEGORY": "C",
    "CITYNAME": "New York",
    "POSTALCODE": "10007"
    }
]}

The HTTP response will looks as follows :

"{ "result": [30.966167923963926, 70.18799357457902]}"

Note : if you're using Postman, you can find a sample test at Postman Sample Test

You can now proceed with the next step.

[Optional] Automated ML - Best Model - Additional Info

If you're interested in the 'Best Model' Auto ML selected, return to your 'Automated ML Run'.

Under Model Summary you can see the Algorithm which was selected as best and the corresponding error.

Under Models, you can see all the algorithms which were evaluated during the Auto ML run.

Select your Best Model, and select View Explanation. Select an Explanation ID and Aggregate Feature Importance.

From this view, you can see that CUSTOMERGROUP was the most important feature to determine the offset.