AI Trend Forecasting

The AI Trend Forecasting module is an optional service designed to forecast labor costs using machine learning. This module enables you to predict future data trends based on historical data, allowing for better labor cost planning and tracking.

The process starts with the creation of a forecast, which is based on historical widget data. The module then uses machine learning algorithms to generate predictions, which are displayed at the end of the process. These algorithms are entirely date-based, meaning that they rely on past performance to predict future performance.

They also take into consideration both country and time-zone, which means holidays and regular work weeks are factored into calculations. This process does not yet accommodate external factors such as major changes in a warehouse’s structure or labor force.

Concepts

Key Definitions

Term	Definition
Forecast	A tool used to analyze historical data through machine learning algorithms.
Pipeline ID	A unique identifier assigned to each forecast run.
ML	Machine Learning

How Trend Forecasting Works

1. A forecast is run, which triggers an API request to the ML application.

2. The ML application processes the request by preparing and pre-processing data, generating results, and storing them in the database.

3. The forecast’s status is updated at each step, keeping you informed about the progress.

4. Finally, the results are generated and include:

• The Execution date & ID

• The Pipeline ID

• The forecasted results of the data you want predicted.

How ML Calculates Predicted Fields

When creating a forecast, you select a widget as the basis of the forecast’s configuration. The widget must have at least one date field and one numerical field, and ideally at least a year of historical data. The ML model will use this data to predict future labor costs.

You have the option of creating a General or an Advanced set-up. The Advanced set-up allows you to filter the widget data you want included in the forecast and use a training model to help the ML application’s accuracy.

After a forecast is run, an API request is sent to a cloud-run instance where the ML application is located. Here it prepares data, completes pre-processing for ML forecasting, generates results, and stores them in the database.

Finally, the forecast results are shown as a grid.

Understanding Forecast Results

Once the results of the forecast are generated, you can view them in a grid that displays the execution date and ID, the pipeline ID, and the forecasted results of the data you wanted predicted. Consistent results require sufficient historical data, preferably a year or more. The less historical data there is, the less likely the predictions will be accurate.

Take this example of someone using Trend Forecasting to predict the Goal Hours for the Receiving task. The task and its historical data are pulled from the widget used to create the forecast. Once the forecast is run and the results are generated, the predicted Goal Hours for the activity can be seen in the results table.

The user in this example can then use this predicted data to make decisions about their future warehouse operations. For example, if a warehouse operator notices a trend that Goal Hours are predicted higher on Fridays, they can adjust the labor scheduled for that day of the week.

Creating Forecasts

Prerequisistes

Before you can create a forecast, you need to have or create a daily Historical Labor or Historical Demand widget. This widget must include:

• a date field

• a numerical field to be predicted

• ideally a year’s worth of historical data

For example, you can set up a widget that contains the daily measured and goal times for a Receiving activity. For more information on how to build a custom widget, see Widget Building.

Note

Once a widget is connected to an active forecast, you cannot modify or deactivate it. You will need to deactivate the active forecast first to make changes.

Steps

1. From the Rebus menu, select Admin Tools >AI Trend Forecasting. You’re at the AI Trend Forecasting screen.

2. Click ‘Create Forecast’.

3. Name your forecast. In the Name field, enter a name for your forecast.

4. Select a dataset to explore.

1. In the Widget field, select a widget from the dropdown menu, or search for the widget you want to use.

2. In the Fields to Predict field, select the numerical fields from the widget that you want to predict.

5. Select a time series.

1. Select a DATE field from the dropdown menu.

2. Select a date range for your future prediction. For instance, to predict the upcoming week, select the beginning and end of the week from the date picker.

3. Select the country and time zone.

6. Click save. You can now activate your forecast, or go to the Advanced tab.

Activating a Forecast

Once a forecast is saved, you can find it on the AI Trend Forecasting screen. Here you can activate your created forecasts and perform four different actions.

• Delete the forecast

• View the forecast results

• Execute the forecast

• Edit an inactive forecast OR

• View the details of an active forecast

To activate and execute a forecast:

1. Click the ‘Inactive’ status toggle next to your forecast name.

2. A confirmation window will pop up. Click ‘Yes’. Once your forecast is active, you cannot make any changes.

3. From here you can execute the forecast. Its status will be updated in the ‘Last Run’ column. The status updates automatically every 15 seconds.

4. Once results are generated, you can view your forecast results. You can also click on the purple link to open Widget Builder directly to create a forecast widget.

Using the Advanced Tab

The Advanced set up allows you to customize ML training and refine your forecasts. Using the Advanced tab means you have specific fields you want to filter, you want to generate multiple predictions of the same data, or you want to experiment with different features to find the most accurate prediction for your data.

Training Partitions

In the Training Partitions section, you can separate data into distinct categories (partitions) to improve prediction accuracy. This is especially useful when different categories of data behave differently and need to be treated separately.

Steps

1. Select ‘Add Partition’.

2. Select the field you want to filter from the dropdown.

3. Select ‘IN’ to include the data and ‘NOT IN’ to exclude it.

4. Add the values you want to include or exclude in the Values section. Separate multiple values with a comma (ex. WH1,WH2).

You have now added a partition. You can add up to 5 total partitions.

Prediction Volume Granularity

In the Prediction Volume Granularity section, you can generate multiple predictions of your data based on work volumes.

For example, you can generate predictions for estimated time spent on Receiving based on a low-volume, medium-volume, or high-volume day. A low-volume day is calculated based on the lowest 33% of historical data, a medium-volume day is calculated using the next 33% of historical data, and a high-volume day is calculated using the highest 33% of historical data. Similarly, you can also predict data with even more granularity, with Very Low, Low, Medium, High, Very High volumes, where historical data is separated in 20% intervals.

Steps

1. Select the quantity field you want to filter from the dropdown.

2. Select the Work Volume values you want predicted (Low, Medium, High or Very Low, Low, Medium, High, Very High).

In your forecast results, you will see multiple predictions for each day, depending on the different volumes selected.

Date Extraction Features

The Date Extraction section contains 6 automatically selected parameters that affect your end data. You must keep at least 3 of these options selected.

Parameter Selected	Result: Trend Forecasting will take into consideration the historical trends of…
Day of the week	different days of the week.
Week of the Month	different weeks of the month.
Holiday	holidays (determined when selecting your country).
Week of the Year	different weeks of the year.
Weekend	weekends.
Leading to Holiday	the two days before a holiday.

Validating a Forecast

Validating your forecast is done by comparing forecast results with actual results. You can then adjust your forecast depending on the accuracy of your prediction.

Steps

1. Create a forecast. When selecting a time series, choose a date that you already have data for.

2. Run the forecast.

3. View the forecast results. Compare the forecast results with the actual data you already have.

4. Evaluate the forecast accuracy. Look for any major discrepancies in the data.

5. Based on the accuracy of the forecast, adjust the parameters set in the advanced tab if needed.

6. Repeat the process.