Fastenal Sales Forecasting

Fastenal Sales
Forecasting
Daniel Morrison
About Fastenal
• Sell industrial parts
• Branches sell parts
• Distribution centers supply branches
Problem Description
• Supplier offers rebate
• 5% rebate at distribution center
• No rebate at branches
• 2.5% carrying cost at end of the month
Problem Formulation
• Predict how many parts will be sold next month
• Validation of forecast
• Simulate forecast predictions to measure performance
• Report prediction to company
Data
• Given Data
• List of all part numbers and price for a distribution center
• Part orders from January 2013 through May 2015
• Date sold, part number, price, quantity
• Parts sold from January 2013 through May 2015
• Date sold, part number, quantity
Data - Cleaned
• Cleaned Data
• The number of a given part sold each month
• The price of each part for a distribution center
• Remove past part orders, branch data, dates,. . .
Linear Regression
• Assume sales are linear
• Very easy
• Doesn’t fit
Modified Linear Regression
• Sales are linear over a few months
• Adds seasonality
• Still not good
Time Series – ARIMA & ETS
• ARIMA – choose how far back to look at data
• ETS – weight data less and less going back
• These are the models we need to make good sales predictions
Validating a Forecast
Validating a Forecast
Validating a Forecast
• 𝑠𝑐𝑜𝑟𝑒 =
2𝑠 − 𝑝 −𝑠
2𝑠
• s = Number of parts sold
• p = Number of parts predicted
Simulation
Jan –May 2015
Run prediction
up to month M
Increment M
Calculate score
up to month M-2
Use sale data to
get rebate & cost
Predict part sales
based on score
Buy parts based
off prediction
Initial Results
• All forecasts were good enough to use
• Threshold of 0
• Profit:
• $150,000 in 2014
• $50,000 in first 5 months of 2015
• Good results, but not quite realistic yet. . .
Extended Problem
• Had been treating the distribution centers as a single location
• Really 17 distribution centers
Master Hub
• Choose where to send parts
• Include transportation costs
Region Hub Region Hub
Primary Hub Primary Hub Primary Hub Primary Hub
Categorizing Parts by Level
• Want to purchase at primary level distribution center
• Reduce transportation cost
• Easy to predict at master hub
• Pick in the middle
• Use prediction score to determine where to order parts
Final Results
• $29,000 in profit in June and July 2015
• Prediction was 65% of optimal
• 70% is highest realistic value
• Single supplier of hundreds
Final Results