DAM 7020 Module 6 Final Project: Forecasting Tire and Brake Demand in R for a Forty-Store Service Chain Example

Forecasting What Drivers Will Need: A Store-Level Demand Model in R for an Auto Service Chain

Student Name

MBA in Data Analytics, Thomas Edison State University

DAM-7020: Predictive Analytics for Business Intelligence

Mentor Name

Month Day, Year

Forecasting What Drivers Will Need: A Store-Level Demand Model in R for an Auto Service Chain

Executive Summary

The chain's current method, using the same month last year, misses monthly demand for tire installations and brake jobs by 13 to 15 percent on average. A regression model built in R, using seasonality, vehicle miles traveled two months earlier, gasoline prices and the number of cold days, cut that error to 6 to 7 percent when tested on 2024, data the model had not seen. The report recommends producing store-level forecasts each month, feeding them into the warehouse's tire orders and store managers' schedules, and reviewing accuracy quarterly.

The Decision and the Data

Two decisions depend on demand: how many tires of each common size the regional warehouse stocks, and how many technicians each store schedules. The midterm joined two public data sets, vehicle miles traveled from the Federal Highway Administration and regional gasoline prices (U.S. Energy Information Administration, 2024), with the chain's monthly service counts for 2015 through 2024. For the final project I added a third public series, the number of days each month with a minimum temperature below freezing at Newark, a federal weather series (National Oceanic and Atmospheric Administration, 2024), because cold snaps drive tire and battery visits.

Data preparation in R followed the rules set earlier in the course: one script for cleaning and joining, one for modeling and one for output, each documented. The 2020 pandemic months were kept but flagged with an indicator variable, since excluding them would break the monthly series. Store-level data were available for 38 of 40 stores for the full period; two newer stores were forecast from the chain total and their share of sales.

Methods

Hyndman and Athanasopoulos (2021) recommend comparing forecasting methods against simple benchmarks and evaluating them on data not used for fitting. I compared three methods at the chain level first, then applied the best to stores.

Models were fit on 2015 through 2023 and evaluated on the twelve months of 2024 using mean absolute percentage error. R software with forecasting packages handled fitting and evaluation (R Core Team, 2024), and all scripts are attached so the analysis can be rerun.

Results

The regression performed best for both outcomes. Its coefficients make sense: each additional cold day in a month was associated with about 1.6 percent more tire installations that month, and a 1 percent increase in vehicle miles traveled was associated with about 0.8 percent more brake jobs two months later. Gasoline prices added little once miles traveled were included, which suggests their effect works mainly through how much people drive.

Applied to stores, the regression's error averaged 9 percent, higher than at the chain level, as expected for smaller numbers, but still well below the seasonal naive method's 18 percent at the store level. Eleven stores near major highways showed stronger links to miles traveled; stores in denser towns were more seasonal and less affected by driving volume.

Interpretation for Managers

Shmueli and Koppius (2011) insist a predictive model be judged by how well it forecasts; any explanation it offers is a bonus. Here we gain both. The accuracy gain means fewer stockouts and less overtime; the coefficients give store managers reasons they can understand, such as a cold forecast meaning more tire business. Because miles traveled arrive with a two-month lag, they suit the brake forecast well but can only be used for tires through the seasonal and weather components, which come from a weather outlook.

Expected Benefits

If the forecast had been used in 2024, the warehouse would have avoided an estimated 60 percent of stockouts of the ten most common tire sizes, based on comparing forecast-based orders with actual demand week by week. Overtime from demand spikes would have fallen by an estimated $150,000 to $200,000, depending on how closely managers followed the forecast in scheduling.

Recommendations

First, by the second business day of every month, issue three-month forecasts of tire and brake work for every store on the dashboard managers already open daily. Second, base the warehouse's weekly tire orders on the chain forecast split by store and size, with a buffer for the most common sizes. Third, ask store managers to schedule technicians using the forecast and to note when they override it and why. Fourth, review accuracy quarterly and refit the model each January.

The Future of Analytics at the Chain

Module 6 looks ahead. Once forecasts are trusted, the chain can add prescriptive tools that recommend transfers of tires between stores, and it can incorporate appointment bookings from its website, which give a direct early signal of demand. As in the course's Module 2 review, the priority is building the habit of using forecasts, which makes later tools worth adding.

Governance and Ownership

Forecasts will become part of how the chain runs only if someone owns them. The analytics manager will own the model and its monthly refresh; the warehouse manager will own the tire ordering rule that uses the chain forecast; and each store manager will own the decision to follow or override the store forecast, recording reasons for overrides. A quarterly review with the operations director will compare forecast accuracy and override outcomes, so that both the model and managers' judgment improve. When the model is refit each January, any change in its predictors or accuracy will be explained to store managers in plain language before new forecasts are released.

Communicating Uncertainty

Every forecast will be shown with a range, not a single number, so managers understand that a forecast of 120 brake jobs means roughly 110 to 130. Hyndman and Athanasopoulos (2021) stress that prediction intervals are as important as point forecasts, because they show how much confidence a forecast deserves. Ranges also help managers decide how much buffer stock to carry for the most common tire sizes.

Training Store Managers to Use Forecasts

Before the first forecasts go out, each store manager gets an hour of instruction. It will cover how to read the forecast and its range, how the forecast connects to the warehouse order, when to override, for example when a local fleet customer has scheduled a large job, and how to record the reason. Managers of the eleven highway stores will learn that their demand follows driving volume more closely, so their forecasts may swing more with seasonal travel. After three months, a short survey will ask managers whether the forecasts are useful and what would make them more so, and those answers will guide the next refit of the model.

Limitations

The model rests on ten years of history, including a pandemic year; another shock could break the relationships. Public driving data are revised after release. The benefit estimates assume managers act on forecasts, which the override log will test. And the model forecasts total demand, not demand turned away when a size is out of stock, so true demand may be slightly higher than recorded.

Appendix: The R Scripts

Script 1 downloads the federal driving, fuel and weather series, aligns them by month and joins them to the chain's service counts. Script 2 fits the three models on 2015 through 2023 and computes 2024 accuracy. Script 3 refits the chosen model by store and writes forecasts to a file the dashboard reads. Script 4 draws the charts used in this report. Each script begins with a description of its inputs and outputs and checks for missing months.

References

Federal Highway Administration. (2024). Traffic volume trends. U.S. Department of Transportation. https://www.fhwa.dot.gov/policyinformation/travel_monitoring/tvt.cfm

Hyndman, R. J., & Athanasopoulos, G. (2021). Forecasting: Principles and practice (3rd ed.). OTexts. https://otexts.com/fpp3/

National Oceanic and Atmospheric Administration. (2024). Climate data online. National Centers for Environmental Information. https://www.ncei.noaa.gov/cdo-web/

R Core Team. (2024). R: A language and environment for statistical computing. R Foundation for Statistical Computing. https://www.R-project.org/

Shmueli, G., & Koppius, O. R. (2011). Predictive analytics in information systems research. MIS Quarterly, 35(3), 553-572. https://doi.org/10.2307/23042796

U.S. Energy Information Administration. (2024). Weekly retail gasoline and diesel prices. https://www.eia.gov/petroleum/gasdiesel/