New Product = Gambling? Forecasting

Transcription

1 Universität Hamburg Institut für Wirtschaftsinformatik Prof. Dr. D.B. Preßmar Predictive Analytics for New Product Forecasting More Analytics Statistics and less Judgment: a Case Study Dr. Sven F. Crone Dr. Nikolaos Kourentzes New Product = Gambling? Forecasting Page 1

2 Both? Statistics? e.g. Norton-Bass Classification Clustering Social-media analysis Judgment? e.g. Delphi Szenarios Market Tests Prediction Markets Statistics vs. Judgment in New Product Forecasting? How to forecast high-tech products (CPU,RAM) at Intel? Product sourcing & production lead times are longer than sales period every product is new Often requires building new production facilities (or even plants) high costs for under & overforecasting s to 1,s of new products per season 2-4 innovation cycles (selling seasons) per year Requires semi-automated & standardized process How to they do New product forecasting? The Challenge Page 2

3 Norton-Bass Model Sales of products with continuous repeat purchasing? Estimate total market volume from first 3 sales observations Fashion Retailers order from China / India 3-6 months lead time 3-12 seasons per year Common Problem Catalogue Retailers order from China / India 3-6 months lead time 4 catalogues per year Sourcing lead times are longer than sales period (for many products) many products are new Page 3

4 Common Problem CDs, DVDs, BlueRays, Video Games Product launch & intial sales before, during & after golive (i.e. pipeline-filling) are most important Common Problem check out the scaled thumbnails of the first 8 weeks of sales time series plots for new DVD releases Page 4

5 Few things are completely new! Forecast using Analogies Delphi Szenarios Market Tests Prediction Markets How to find similar products from history (aka Analogies) automatically? Finding Analogies: you try it! aka statistical technique called Clustering Cluster 1: Cluster 2: Now find: Page 5

6 Analogies in Time Series? Hard to extract useful information! No additional information Find time series that behave similarly using Kmeans clustering Focus on the first three months (new products) Time series with 27 months of observations Finds clusters of typical initial sales behavior Analogies in Time Series? New product to be launched Multiple Only one observations initial order (plus 1s of product attributes etc.) Match to most similar cluster Compare similarity Take average shape of cluster as forecast! Need to be careful in specifying SIMILIARITY! Page 6

7 What is Similarity? What is Similarity? Page 7

8 What is Similarity? What is Similarity? Page 8

9 What is Similarity? What is Similarity? Page 9

10 Similarity in Time Series? Euclidean Distance between two time series Q = {q 1, q 2,, q n } and S = {s 1, s 2,, s n } Q Database Distance.98 Rank 4 n, D Q S q i s i i 1 S 2 Find similar shapes in time series D Q, C p n q i c i i 1 p.43 3 n datapoints Max (p=inf) Manhattan (p=1) Weighted Euclidean Mahalanobis How do we know that it works? Frost! Theory? Not a good idea Empirical Testing! Tongue! Page 1

11 Curtains Tiebacks Cushion covers Valances in various fabrics, linings, sizes, colours Case Study Task(s) Clustering Approach allocate to clusters on more data Clustering of the 1 st pre-order & product attributes Clustering on the first three observations Objectives? support complete launch process! Forcast inital pipline filling Without sales history Initial behaviour matching Very short history Clustering on the first six observations Longer history Recalibrate matching after the 6 th month the history is long enough to use time series methods Page 11

12 Sample new product First 52 weeks of deliveries Pipeline filling Normal behaviour Know initial delivery volume How to forecast with zero weeks of sales? Stage I - Clustering Products Stage II - Forecast X 3 units 75 units 45 units... Initial known orders (interview, trial etc.) Sales history Classification to repository New Product Clustering time series clustering Adapt pattern to new product Repository of clusters (with archetypical new product sales) Calculate safety stock and use! Automatic forecasting by structured analogy Page 12

13 Clustering Results 3 first observations I. 262 products used with history up to 27 months II. Each product Only first three months used III. 8 clusters Optimal number IV. Each cluster Average behaviour found V. Average behaviour New products will be matched with Clustering Results 3 first observations Cluster Cluster 1 Members: 1 14 Cluster Cluster 2 Members: 2 22 Cluster Cluster 3 Members: 3 56 Cluster Cluster 1 Members: Cluster Cluster 2 Members: 5 9 Cluster Cluster 3 Members: 6 18 Cluster Cluster 4 Members: 7 3 Cluster Cluster 5 Members: Business intelligence can eliminate/correct clusters Cluster 5 case Page 13

14 Forecasting I. Using the estimation for the first month value and the product type find the correct cluster using the three first observations clustering e.g. First value = 15, Unlined curtain Cluster 7! Forecasting I. II. Using the estimation for the first month value and the product type find the correct cluster using the three first observations clustering Use the % change of the average shape of the cluster to form the 2 nd and 3 rd value Second Value (Month): 15 * (1-,4823) = Third Value (Month): * (1+,244) = Page 14

15 Forecasting I. II. Using the estimation for the first month value and the product type find the correct cluster using the three first observations clustering. Use the % change of the average shape of the cluster to form the 2 nd and 3 rd value. For the 4 th -6 th month forecasts use the clustering of the first 6 months III. to find the product membership (similar procedure!). Update the forecasted values with real whenever they are available. IV. If possible re-evaluate membership when new values are available. V. To forecast after the sixth value/month use a statistical model. The product history is now long enough! New Product Forecasting Experimental Evaluation How accurate is it? Evaluate on time series of the given data Proposed Approach 1 8 Test1 Test2 Test3 Forecast1 Forecast2 Forecast3 Statistical Forecasting Time series clustering always outperforms statistical approach Time series clustering outperformed human demand planner Page 15

16 New Product Clustering Allows us to measure safety stock! Historical Data Clustering Profiles Inventory Performance Results Empirical Evaluation Inventory Performance Clustering on 16 Sales 16 Sales & Features S1: Features S2: Initial Orders Stock-out 8 Stock-out 8 S3: Features & I. Orders 6 6 S4: Features & Sales (2) Inventory Inventory S5: Features & Sales (6) Page 16

17 Clustering Results A simple Forecasting Tool Outcome: simple MS EXCEL application that can be automated Take aways Looking for companies for a free-of-charge pilot study (to present at IBF?) Many industries specialize on New Product Forecasting opportunities to learn (steal fire) Clustering creates insightful groups of similar products helpful to planners to improve judgment Allows large-scale automation of new product forecasting Predictive Analytics provides new solutions to forecasting challenges room for improvement! [LCF ISF9, SAS New Product Presentation ISF 8] Page 17

18 Questions? Dr. Sven F. Crone Assistant Professor, Director Lancaster University Management School Research Centre for Forecasting Lancaster, LA1 4YX Tel Copyright All rights reserved, Dr. Sven F. Crone These slides are from a presentation on forecasting. You may use these slides for internal purpose only, so long as they are clearly identified as being created and copyrighted by Sven F. Crone. All rights are reserved! You may not use the text and images in a paper, tutorial or external training, duplicate or replicate or alter them, or make them accessible as is, without explicit prior permission from the author. Dr. Sven F. Crone Assistant Professor, Director Lancaster University Management School Research Centre for Forecasting Lancaster, LA1 4YX Tel s.crone@lancaster.ac.uk Page 18