M-Factor M3 Aggregates Segment-Level Mix Models (Which Is Cooler Than It Sounds)

I spend most of my time these days thinking about business to business marketing, where performance is measured one sale at a time. Ironically, it’s much harder for business marketers to calculate their impact on sales than for marketers in the anonymous, vastly less precise realm of consumer packaged goods. The reason: predicting individual behavior is difficult in both cases, but there are so many consumers that their aggregate behavior can be modeled accurately with statistics. Packaged goods marketers may not know the name of every tree, but they have a much clearer picture of the forest.

The main tool used to measure consumer marketing results is, of course, the marketing mix model. This is built by identifying historical correlations between sales results and inputs such as media spend, trade promotions, pricing, primary demand and competitive activities. Mix models can provide powerful insights into the causes of past performance and helpful forecasts of the impact of future plans. Even though few marketing managers really understand the underlying math, the models are well enough proven to be widely accepted.

But there are limits to what a single marketing mix model can accomplish. Most markets are in fact comprised of many different segments, based on geography, customer type, product attributes, and other distinctions. Each segment will behave slightly differently, so generating the most accurate results requires a separate model for every one. This wouldn’t matter, except that marketers work at the segment level. They have separate marketing plans for each segment and track segment results. In fact, a large company will often have entirely different people responsible for different segments. You can be sure that each of them focuses on her own concerns.

Building lots of segment-level models doesn’t have to be much more expensive than building one big model. The trick is keeping the inputs and model structure the same. But managing all those models and aggregating their results does require a substantial infrastructure. This is what SAS for Marketing Mix (formerly Veridiem) was designed to do. (See my related post ) . It’s also the function of M-Factor M3.

In fact, M-Factor was originally founded in 2003 specifically to help combine marketing mix models that were created by third parties. The company’s product can do this, but the firm found that externally-built models are often poorly understood, difficult to maintain, and inconsistent with each other. In self-defense, it decided to build its own.

Today, M-Factor developed its own model-building staff and toolkit. This allows it to develop separate models for each segment in a market—sometimes hundreds or thousands of them. These can be arrayed in a multi-dimensional cube, which allows the system to easily aggregate results or drill down within different dimensions. Sharing the same structure also makes it easy to update the models with new data and to build detailed reports such as profit statements derived from model outputs.

To go at it a bit more systematically, M3 provides three main functions. The first is results analysis: calculating return on marketing investments by estimating the contribution of each input to over-all results. The second is forecasting: accepting scenarios with planned inputs, and using these to estimate future results. The third is optimization: automatically identifying the best combination of inputs to produce the desired outputs.

The results analysis accepts historical inputs from the usual sources such as Nielsen and IRI. It then produces typical marketing mix reports on the sales levels, volume drivers and return on investment. It also provides model performance reports such as model fit and error analyses. M-Factor makes a point of breaking out the model error, to help users understand the limits of model accuracy and see how well models hold up over time. The company says that its particular techniques make its models unusually robust.

Forecasting starts with a marketing plan for business inputs such as budgets and prices. These are at roughly the same level as the mix model inputs: that is, spending by category but not for specific marketing campaigns. A typical model has 15-25 such inputs. They can be entered for individual segments and then aggregated by the system, or the user can provide summary figures and let the system distribute them among segments according to user-specified rules. The system then applies these inputs to its models to generate a forecast.

Once an initial plan is entered, it serves as a base for other scenarios. M3 displays the original inputs as one column in a grid, and lets users make changes in an adjacent column. Since the models are already built, the forecast is calculated almost instantly. Results can include a full profit statement as well as the inputs and estimated sales volume.

Users can freeze one forecast to treat it as the business plan. The system can later report planned vs. actual results, or compare the original plan against a revised forecast. The system can also project results for the current calendar year by combining actuals to date with forecasts for the balance of the period. Because the forecasts are built by the individual segment models, all results can be analyzed via drill-downs or aggregated into user-defined groups. M3 provides each user with a personalized dashboard to make this easier.

Optimization is an automated version of the scenario testing process. The user specifies output constraints such as minimum revenue levels, and driver ranges such as no more than 3% price change. The actual optimization process uses a genetic algorithm that randomly tests different combinations of inputs, selects the sets with the best outcomes, makes small changes, and tests them again. It continues testing and tweaking until it stops finding improvements.

Users can also ask the system to optimize two target variables simultaneously. What the system actually does is combine them into a weighted composite, using different weights in different model runs. It plots the result of each run on a chart where the X axis represents one target variable and the Y axis represents the other. Users can then choose the balance they prefer.

Initial deployment of M3 usually takes three to four months, including the time to assemble the historical data, build the models, and provide an initial set of strategic recommendations. Pricing is comparable to conventional mix models, although it is sold as a hosted service on an annual subscription. This typically includes monthly data updates and reports, and quarterly updates of the underlying models. End-users access the system via a browser and can run reports, scenarios and optimizations at will.