Forecasting and Yield Optimization

Large scale publishers, ad networks and similar such online inventory aggregators have stepped up their quest for yield optimization and revenue maximization. I've seen this before, so allow me step back in time...way back (in Internet terms) to the year 1998. I was a few years out of college. After a brief stint in the Finance Industry working on a trading desk at a broker dealer, I decided that my future lay in building businesses. I took a position in inventory and revenue forecasting at the Vermont Teddy Bear Company where I learned how to program and build databases. Shortly thereafter I was recruited by Sara Lee-Champion Jogbra where I really dug into the statistical aspect of forecasting. In 1998 statistical forecasting techniques were just beginning to get attention as a potential profit optimizer for the company. Using some data mining techniques and the statistical forecasting engine, Forecast Pro, I realized we were making the wrong ratio of bra sizes, particularly when you look at the size ratio by style, for instance a more supportive bra with an underwire ran to the larger sizes, where a compression style ran to the smaller sizes (yes I am an expert in both teddy bears and sports bras.) Given that sports bra design and colors for a segment of the product line were seasonal decreased the lead time for making adjustments to production. Additionally, unsold seasonal sports bras qualified for return to manufacturer where we would reroute them to bargain outlets thus hitting our margins (ahhh yes...the direct correlation between inventory forecasting and revenue forecasting.) We had to get it as right and as real time as possible. With a new forecasting strategy I was able to predict the size ratio demands for the different categories of inventory using exponential smoothing (essentially an algorithm of regression with seasonality adjustments) for styles with a longer data set and weighted rolling averages for newer styles. I took that further in future iterations, by comparing a new style that had an affinity with older styles and adding an adjustment into the rolling average weight, and I extended the forecasting process to include feedback from field sales for making macro level adjustments on demand. At each stage I would do post-hoc testing to validate the algorithm changes using experimentation against known results. The net result was significant reduction in returned bras, and a big increase in sales because prior to the project we were selling out of the correct sizes.

So how does this relate to inventory optimization in the online environment? Everything...think about the correlations; forecasting at higher grains, field sales integration, data mining, algorithmic strategy and modeling, post-hoc testing, simulation, ephemeral inventory, seasonality, inventory performance affinities, margin hits on channel routing to remnant etc...

Like 1998 in manufacturing the online industry is rapidly innovating on inventory (read "yield") optimization. Sophisticated solutions are popping up along with some rising startups; The Rubicon Project, Yieldex, Pubmatic, AdMeld to name a few. Each with a different flavor for improving yield, some more technically oriented, some more strategically oriented. Throw into the mix multiple channels (and channel conflicts, oh boy fun) for selling the inventory; in-house sales, ad networks, ad exchanges, of which I know a few things ;), and you have an amazing transformation in the industry. Now we just have to figure out how to make it all work together. So let's get back those forecasting books. Here is a refresher on some key concepts:

Forecasting Methods

• Extrapolative Methods - Find a pattern(trend) to the historical time series and assume that the pattern will continue into the future.
• Explanatory Methods - Determine the factors which explain the past behavior of the variables to be forecast.
• Judgmental Methods - Methods which rely, not upon a statistical analysis of the historical data, but on Managerial expertise, feedback from Sales Reps and Customer surveys.

My experience is that you use all three and over time strive to find the right mix of the three and the right mix of possibilities within the three.

Forecasting Indicators

• Leading Indicators - Provide advance notice
• Concurrent Indicators - Provide confirmation at the time of the event
• Lagging Indicators - Provide notice after the event

While it is tempting to dismiss Lagging Indicators, they are useful for post-hoc tests of new econometric models (Extrapolative Methods) due the "Assumption of Constancy" (the patterns of the past indicate the patterns of the future.)

Special thanks to Dr. Len Tashman director of the Institute for Forecasting Education and editor of Foresight - The International Journal of Applied Forecasting for his many years of support and mentoring. Contact the Institute if your company is looking for an evaluation of your revenue and/or inventory forecasting methods.

Fair Use Policy: Please include the name: Edward Boettcher and the following link: The Product Messenger when referencing this information. All content (c) 2009 Edward Boettcher