Sale forecasting to reduce food discard

Sale forecasting to reduce food discard#

Customer pain-points#

According to estimates by the Food and Agriculture Organization (FAO) of the United Nations, the cumulative amount of food lost or wasted globally reaches 1.3 billion tons per year, which is equivalent to 1.6 times the total amount of food donated to famine victims. According to the U.S. Department of Agriculture (USDA), between 30% and 40% of food is wasted in the U.S., with 31% of this occurring at the retail and consumer level. The mismatch between supply and demand is one of the main causes of food loss. Therefore, we expect to reduce the discards by accurately forecasting the next days of sales.

Solution effects#

Collaborating with a large retailer, we are able to improve unnecessary waste to about 1.3% of total sales. As a reference, the EBITDA in the fiscal year of this retailer is 6%.

../_images/forecast_reduce_discard_illustration.png ../_images/forecast_backtest.png

Behind technology#

Historical customer behaviors drives sales. Understanding relations among different behaviors as well as their time-elapsed effect are the key to business success. Unfortunately, time-indexed dataset can be treated multiple viewpoint. For example, the five samples of ’Order’ can be understood as five independent observations of the random variable ‘Order’, or one observation of a random process ‘Order(t)’. Moreover, which variables (e.g., browse only, or with price) and/or what is the lagged information of a specific variable (e.g., browse a day ago) should be considered often yield different results.

../_images/confusion_to_decision_maker.png

In this project, you input a 2D spreadsheet where there is a column to represent timestamp. We firstly pass this time-series data a set of rigorous statistical test (e.g., Ljung-Box and granger causal test) and transform your data to be prepared with mathematical assumptions. Then, we auto-determine N-th Markovian to be a sufficient patterns, fitting the model, and providing a suite of temporal model analysis. It contains the following functions:

  • “What will happen next” yields multi-step forecasting with confidence bonds.

  • “Why this happen” tells one what are the cause factors to a given target variable.

  • “Strategy simulation” imagines different scenarios prior actual execution.

../_images/highlight_mts.png

Reference

  1. US20220414540 patent. (pdf)

  2. US20240070160 patent. (pdf)

  3. Erli Wang, Causal Analysis with Application to Inventory Control. DataFunSummit 2023. (link)

Contents