What are the demand forecasting models commonly used in industry when dealing with substitutable products under cross price elasticities?
Often sellers offer a portfolio of substitutable products. Amongst other factors, an individual product's demand could depend on not only the price kept for the product, but also on the prices for the other products in the seller's kitty. In such situation, how can good demand estimates be made?
Consider for example a car dealership. The dealership is carrying a variety of car models. When a customer walks into the dealership, he is likely to select one of the models available at the distributor. His decision in selecting a specific car model is not only effected by the price of that car model, but also the prices for the other car models available at the distributor. In such situation, how can the dealership make a good estimate of the demand for different car models?
Choice models can address these situations, but they are restrictive due to the strong underlying assumptions. I would appreciate if someone can point me to the industry practices in this regards.
Thanks in advance
Anshu Jalora
Good Answers (4)
Venkatesh R.
Consultant, independent researcher, blogger (http://ribbonfarm.com), author of "Tempo" (http://tempobook.com)
Best Answers in: Business Analytics (3), Organizational Development (2), Writing and Editing (1), Career Management (1), Enterprise Software (1)
Ugh. It is a very difficult problem. I'd be suspicious of any first principles models no matter how compelling their underlying logic. You should try look up ARMA and ARIMA models (described in any book on time-series analysis). In general, for market-spanning portfolios, you are talking about mature markets where there will also be competitors, so you really need to talk about demand distributions among segments and product portfolio at once, and add in business cycle effects. In other words, it is a complicated mess. So first get a sense of how your real data behaves with the models above, then you can drill down and build more tasteful models.
I recommend profitics.com to find expertise in this domain (objective advice only if you ignore the fact that I am the CEO of profitics).
Demand planning processes which include demand forecasting are business processes and not just mathematical models. Enterprise software that enables the processes described on profitics.com is the most important part of solving that problem. Regarding mathematical models, domain specific variants of ARIMA models are there most popular. Depending solely on historical data and mathematical models for this problem is like driving a car solely by looking in a sophisticated rear view mirror.
For the specific question you asked about the automotive industry problem, the detailed solution is a little longer than this forum would allow, please feel free to contact profitics.
Links:
Boris R.
Managing partner at Bright Capital
Best Answers in: Offshoring and Outsourcing (1), Inventory Management (1), Supply Chain Management (1)
Anshu, based on what I've seen in retail:
- if you can assume for a selected merchandise to have a stable positioning (i.e, car dealership is not going to change relative price levels of different cars), you can make a straightforward forecast using whatever method like ARIMA etc
- if you can not, then there are two options:
- if substitution is close to perfect (in our case, generic drugs are perfect substitutors) - then you have to:
- model total customer expenditure for a specific need rather than individual item - using any forecasting model
- assume some breakdown based on individual sellers forecast, manufacturer, pricing levels etc
- If substitution is not perfect, you'll have to find driving variable and use some kind of regression model, as it was suggested in the previous answer
Anyways, the simplier the better.
Regards,
Boris
You are getting good advice here. The problem as others have cited is bipartite: first, you need to understand existing behaviour and second you need to model future behaviour.
I have been working on a solution for the former problem in retail banking where product substitution on the deposits side of the business is very significant, because money is the ultimate perfect substitute. Being able to understand the substitution effects between different flavours of checking, saving, term deposit and investment accounts is critical in this industry.
We have developed a measurement approach that can be applied to any industry to identify historical behaviour explicitly and lay down time series events that show new business, lost business and internal substitution effects.
Having this data generated in time series enables a bank to:
1. Clarify the target behaviour definition used in modelling by removing false positives and
2. Use the behavioural events as predictive variables.
We don't do the modelling, we only do the metrics side. If you want to know more about this method please see the link below or contact me directly.
Disclosure: I am the inventor of this method and have financial interest in the company linked below.
I hope this is somewhat helpful to you. Good luck !
Links:
More Answers (2)
Rick F.
Supply Chain Management & Manufacturing
Best Answers in: Supply Chain Management (37), Inventory Management (19), Manufacturing (14), Small Business (7), Purchasing (3), Packaging and Labeling (3), Project Management (3), Quality Management and Standards (3), Distribution (3), Starting Up (3), Enterprise Software (2), Facilities Management (1), Education and Schools (1), Economics (1), Exporting/Importing (1), Customer Relationship Management (1), Business Analytics (1), Planning (1), Business Plans (1), Incorporation (1), Computers and Software (1), Wireless (1)
Anshu, this question causes me to have flash-backs to grad school! Try the links I've provided. Best of luck & I hope this helps Anshu.
Links:
Hi Anshu,
In my earlier assignment,I faced almost every month the same issue.I was working earlier as a Demand Planner for a Pharma Company in India.What I did was use a Multiple Variable Regression Analysis as my Forecasting Model.Parameters used were length of Production in days of the particular SKU (specially if you run production in campaigns),Price etc.We got a reasonably high R-square value and MAPE improved to a level of 15%.
The trick in this case is in selecting the right parameters for your model.In our case,we did a trial and error approach.
Hope this helps.
Regards,
Arijit
