Improving Forecasts for Gift Card Sales

The Challenge

A leading food service chain hired Elder Research to improve on a manual forecasting process to project monthly sales from retail and mobile gift cards for annual budgeting and to better-understand quarterly budget variance.

The Solution

Elder Research partnered closely with client subject matter experts to develop and deploy efficient and accurate time-series models to streamline and improve forecasts of future gift card sales related to two sales sources, one of which was related to the client’s loyalty app.

During the data discovery phase, we explored available historical data to understand its coverage, seasonality and overall trends, and prepared the data for modeling. In the case of one gift card sales source where insufficient data existed to accurately forecast strong seasonal trends, we used data from a third source that showed significant similarities in seasonality and data scale to the source of interest. Insights about the data were shared with the client during data discovery, giving the client valuable information and recommendations on the trends and opportunities for improvement.

In the modeling phase, we assessed time-series forecasting algorithms such as ARIMA, Exponential Smoothing, and Prophet to understand each model’s accuracy relative to the client’s baseline forecasts and actual sales observed during the forecast testing window. A robust cross-validation framework was developed to assess the forecast’s accuracy across multiple time periods and forecast horizons. Planning for deployment, we also incorporated our understanding of the client’s local and global environment to assess each model’s ease-of-use and potential barriers to end-user adoption. A fine-tuned version of the Prophet algorithm, a highly tunable method designed for business forecasting, delivered the best results in the case of both gift card sales sources. Prophet incorporates seasonality, holiday effects, and changing trends over time to produce very consistent forecasting results.

In the deployment phase, we created an Alteryx workflow to automate the forecasting process and incorporated the new forecasts into a suite of Tableau dashboards that allows the subject matter experts to easily view forecasts and actual sales with the most recent data available. Budget items are also calculated based on the forecast, streamlining budget planning decisions. The solution allows end-users to make manual adjustments to the forecast to account for additional data that was unknown at the time of budgeting. Examples include the addition of new gift card sales locations or non-regular promotions. Visualizations were built into the dashboards to help the subject matter expert and their internal stakeholders understand the expected error of the forecasts and when to investigate if the model should be refreshed.

MAPE by Forecast Horizon

Results

The final deliverables for this client included the insights identified in their existing data, Tableau dashboards built to their specification, and documentation on how the model was derived and how to use the Alteryx and Tableau interfaces. The solution simplifies the annual budgeting process and provides accurate tracking and evaluation of forecasts at monthly and quarterly intervals. A forecast process that used to take several hours each month can now be completed in less than a minute, with an interface that is integrated with the existing environment. Hands-on onsite training enabled seamless transfer of model ownership to the subject matter experts and strengthened forecasting capabilities within the client’s team.

Budget Variance