•

1 like•1,500 views

Predicting the future is hard and it requires a lot of assumptions, also known as beliefs, also known as faith. In “Assumptions: Check yo self, before you wreck yo self” we explore the consequences of beliefs when constructing predictive models. We’ll walk through the process of developing a demand forecast for Evo, a Seattle-based outdoor recreation retailer, and discuss how assumptions influence the behavior of your application and ultimately the decisions you make.

- 1. Assumptions: Check yo self, before you wreck yo self. Erin Shellman @erinshellman Seattle Software Craftsmanship August 28, 2014 !
- 2. Assumptions: Making an ass out of you and me. Erin Shellman @erinshellman Seattle Software Craftsmanship August 28, 2014 !
- 3. I’m Erin, and I’m a data scientist.
- 4. How much should this cost?
- 6. …and when? What about these?
- 8. Price optimization 1. Git yer Big Data!
- 9. Price optimization 1. Git yer Big Data! 2. Forecast demand
- 10. Price optimization 1. Git yer Big Data! 2. Forecast demand 3. Optimize price
- 11. Price optimization 4. Profit!!!!! 1. Big Data! 2. demand 3. price
- 12. Price optimization 1. Git yer Big Data! 2. Forecast demand 3. Optimize price max X yi = !0 + !1xi + ✏i revenue
- 13. The key is a good forecast.
- 15. Do the easiest thing •Subset the data and focus on one category of product. • e.g. Alpine ski bindings. • Prototype & validate in R. Units Soldi = α + β1(pricei) + εi
- 16. Do the easiest thing •Subset the data and focus on one category of product. • e.g. Alpine ski bindings. • Prototype & validate in R. Units Soldi = α + β1(pricei) + εi Residual
- 17. Assumptions of SLR •We assume that residuals: 1.Normal, with mean zero. 2.Are not autocorrelated. 3.Are unrelated to the predictors.
- 18. Checking assumptions is hard •…and boring! •For statistical methods, assumption testing traditionally relies on visually inspecting plots (and lets be real, most people don’t even do that).
- 19. 40 60 80 100 120 0 500 1000 1500 2000 2500 Fitted values Residuals Residuals vs Fitted 117914 156 -3 -2 -1 0 1 2 3 0 2 4 6 8 Theoretical Quantiles Standardized residuals Normal Q-Q 194 171 156 40 60 80 100 120 0.0 0.5 1.0 1.5 2.0 2.5 Fitted values Standardized residuals Scale-Location 117914 156 0.00 0.01 0.02 0.03 0.04 0 2 4 6 8 Leverage Standardized residuals Cook's distance 1 0.5 Residuals vs Leverage 119741 109
- 20. OF all the practices you can leverage to assist your craftsmanship, you will get the most benefit from testing. ! Stephen Vance
- 21. test_that assumption! context("Check assumptions of SLR") ! test_that("The residuals are normally distributed", { ! expect_that(shapiro.test(model_object$residuals)$p.value, is_more_than(0.05)) ! }) ! test_that("There is no autocorrelation", { ! expect_that(lmtest::bgtest(model_object)$p.value, is_more_than(0.05)) ! }) ! test_that("The residuals are unrelated to the predictor", { ! expect_that(cor(model_object$residuals, data$covariates), equals(0)) ! }) !
- 22. Tests pass! > test_file("./tests/test_slr.R") Check assumptions of SLR : [1] "units_sold ~ price" ... !
- 23. Psych. > test_file("./tests/test_slr.R") Check assumptions of SLR : [1] "units_sold ~ price" 1.. !! 1. Failure(@test_slr.R#12): The residuals are normally distributed ------------------------ shapiro.test(model_object$residuals)$p.value not more than 0.05. Difference: 0.05 !
- 24. Linear? Eh. •We assumed the 2500 functional form was 2000 linear, but there are 1500 several common forms 1000 that might better fit the 500 data. 0 100 200 300 400 500 Price ($) Units Sold
- 25. Price ($) Units Sold Price ($) Units Sold Price ($) Units Sold Price ($) Units Sold Linear Log-log Linear-log Log-linear
- 26. Price ($) Units Sold Price ($) Units Sold Price ($) Units Sold Price ($) Units Sold Linear response to change in price. Much more sensitive to change in price. More gradual response to changes in price Sensitive initially, then gradual
- 30. # Automagically explore SLR with common functional forms candidate_models = list(linear = 'units_sold ~ price', loglog = 'log(units_sold + 1) ~ log(price + 1)', linearlog = 'units_sold ~ log(price + 1)', loglinear = 'log(units_sold + 1) ~ price') ! run = function(candidate_models, input_data) { forecasts = list() test_input = data.frame(price = 0:1000) ! # Forecast for (model in candidate_models) { test_environment = new.env() ! # Generate the forecast forecasts[[model]] = generate_forecast(model, input_data) ! # Save off current value of things for testing assign("model", forecasts[[model]], envir = test_environment) assign("errors", forecasts[[model]]$residuals, envir = test_environment) assign("covariate", input_data$price, envir = test_environment) assign("label", model, envir = test_environment) ! save(test_environment, file = 'env_to_test.Rda') ! # Run assumption tests test_file("./tests/test_slr.R") ! #### OPTIMIZE PRICE!!! #### opt_results = optimizer(forecasts[[model]], test_input) ! # Multiply the predicted demand by the price for expected revenue opt_results$expected_revenue = test_data$price * opt_results$predicted_units_sold ! pdf(paste(model, “.pdf”, sep = ‘’)) plot_price(opt_results) ! } ! return(forecasts) ! }
- 31. rut roh… > run(candidate_models, slr_data) Check assumptions of SLR : [1] "units_sold ~ price" 1.. !! 1. Failure(@test_slr.R#12): The residuals are normally distributed --------------------------------- shapiro.test(linear$residuals)$p.value not more than 0.05. Difference: 0.05 ! Check assumptions of SLR : [1] "log(units_sold + 1) ~ log(price + 1)" 1.2 !! 1. Failure(@test_slr.R#12): The residuals are normally distributed --------------------------------- shapiro.test(linear$residuals)$p.value not more than 0.05. Difference: 0.05 ! 2. Failure(@test_slr.R#24): The residuals are unrelated to the predictor --------------------------- cor(test_environment$errors, test_environment$covariate) not equal to 0 Mean absolute difference: 0.05545615 ! Check assumptions of SLR : [1] "units_sold ~ log(price + 1)" 1.2 !! 1. Failure(@test_slr.R#12): The residuals are normally distributed --------------------------------- shapiro.test(linear$residuals)$p.value not more than 0.05. Difference: 0.05 ! 2. Failure(@test_slr.R#24): The residuals are unrelated to the predictor --------------------------- cor(test_environment$errors, test_environment$covariate) not equal to 0 Mean absolute difference: 0.04201906 ! Check assumptions of SLR : [1] "log(units_sold + 1) ~ price" 1.. !! 1. Failure(@test_slr.R#12): The residuals are normally distributed --------------------------------- shapiro.test(linear$residuals)$p.value not more than 0.05. Difference: 0.05
- 32. 20000 15000 10000 5000 0 Linear Log-log 0 250 500 750 1000 Price ($) Expected Revenue 15000 10000 5000 0 0 250 500 750 1000 Price ($) Expected Revenue Linear-log Log-linear 6000 4000 2000 0 0 250 500 750 1000 Price ($) Expected Revenue 60000 40000 20000 0 0 250 500 750 1000 Price ($) Expected Revenue
- 33. 20000 15000 10000 5000 Optimal Price = $322 0 Linear Log-log 0 250 500 750 1000 Price ($) Expected Revenue 15000 10000 5000 0 0 250 500 750 1000 Price ($) Expected Revenue Linear-log Log-linear 6000 4000 2000 0 0 250 500 750 1000 Price ($) Expected Revenue 60000 40000 20000 0 0 250 500 750 1000 Price ($) Expected Revenue Optimal Price > $1000 Optimal Price = $∞ Optimal Price = $779
- 35. Mean = 185 40 30 20 10 0 100 200 300 400 Price ($) Counts
- 36. In conclusion, these forecasts suck. We are just getting warmed up!
- 37. Beginner-Intermediate Intermediate-Advanced Advanced-Expert 2000 1500 1000 500 0 0 100 200 300 400 5000 100 200 300 400 5000 100 200 300 400 500 Price ($) Units Sold
- 38. 2011-06-01 2011-10-01 2012-02-01 2012-06-01 2012-10-01 2013-02-01 2013-06-01 2013-10-01 2014-02-01 Date Units Sold
- 39. 2011-06-01 2011-10-01 2012-02-01 2012-06-01 2012-10-01 2013-02-01 2013-06-01 2013-10-01 2014-02-01 Date Units Sold TIME?!
- 40. Try something a little smarter Units Soldi = α + β1(pricei) + β2(abilityi) + β3(monthi) + εi
- 41. Beginner-Intermediate Intermediate-Advanced Advanced-Expert 15000 10000 5000 0 15000 10000 5000 0 15000 10000 5000 0 15000 10000 5000 0 15000 10000 5000 0 15000 10000 5000 0 15000 10000 5000 0 15000 10000 5000 0 15000 10000 5000 0 15000 10000 5000 0 15000 10000 5000 0 15000 10000 5000 0 1 2 3 4 5 6 7 8 9 10 11 12 0 250 500 750 1000 0 250 500 750 1000 0 250 500 750 1000 Price ($) Expected Revenue
- 42. Yeah, but who cares? •Do we need to throw everything out just because some assumptions are invalidated? •What is our goal? •Is it still better than what we did previously?
- 43. Wrap it up. 1. Do the easiest thing first, and do it well. It’s how you’re going to learn the domain, and it’s your benchmark for improvement. 2. Test your assumptions, and invest time in building the tools needed to do that effectively. 3. Be cool, stay in school.
- 44. Thanks bros!! Nathan Decker, Brian Pratt & the Evo crew Jason Gowans & Bryan Mayer Elissa “Downtown” Brown, forecasting genius John Foreman, MailChimp #nordstromdatalab
- 45. Click-bait! 1. Data Carpentry: http://mimno.infosci.cornell.edu/b/articles/carpentry/ 2. Getting started with testthat. http://journal.r-project.org/archive/2011-1/ RJournal_2011-1_Wickham.pdf 3. Clean Code: http://www.amazon.com/Clean-Code-Handbook-Software- Craftsmanship/dp/0132350882/ 4. Quality Code: http://www.amazon.com/Quality-Code-Software-Principles- Practices/dp/0321832981 5. Revenue Management: http://www.amazon.com/Practice-Management- International-Operations-Research/dp/0387243763/ 6. Pricing and Revenue Optimization: http://www.amazon.com/Pricing-Revenue- Optimization-Robert-Phillips-ebook/dp/B005JTDOVE/ 7. Original G, Rob Hyndman: https://www.otexts.org/fpp and http:// robjhyndman.com/hyndsight/