We hear a lot about demand forecasting in our industry, and it's an established practice in TV analytics. Ratings forecasting has been a bedrock of TV advertising sales for decades. As the volume and variety of data in the TV industry increases, so does demand forecasting. But what is it, and why is it important?
Simply put, demand forecasting measures past demand to predict future demand. So, if the demand for cars has risen and for motorbikes has fallen this year, demand forecasting would indicate that it is sensible to increase car production and decrease motorcycle production right now to change the supply available next spring.
Sounds simple. However, if this demand forecast had occurred in the fall of 1968, it would have been entirely wrong. The spectacular success the following year of the motorcycle-themed movie Easy Rider, which popularized, and created an unexpected surge in demand for, motorcycles, causing issues in the supply chain. This sudden surge illustrates why demand forecasting is so vulnerable to unforeseen changes between the moment you decide to make the product and the moment where the product is ready to be sold.
Today, the solution to predicting difficult-to-predict surges has shifted from research to mining big data sets to predictive analytics that tries to answer the challenging questions of demand forecasting in a report, which decision-makers then require immediately before making those important decisions which cannot then easily be changed a few months down the line.
Forecasting demand for television
In the television industry, decision-making might be when a broadcaster commits to buying a comedy series, chooses to invest more in live sports events next year, which shows to broadcast or release and when to do so, as well as critically important decisions about inventory pricing. Our data science consulting team carry out demand forecasting projects to help decision-makers answer these questions. So how do data scientists deal with the problem of demand forecasting for the TV industry?
Demand forecasting focusses on all the material being broadcast, ordering shows by popularity and genre, based primarily but not entirely on past audience ratings because audience tastes change over time. Nielsen traditionally dominated these ratings, but nowadays a lot of new data sets are being added into the mix, such as smart TV data and set-top-box data.
However, the goal is not merely to have a comprehensive view of what customers did last spring but to understand what consumer demand is likely to be next spring. If a broadcaster can get a more accurate picture of likely consumer demand in the coming year, this will help them acquire more customers, whereas if they neglect to do so, they will likely lose customers. TV broadcasters want to offer a service that the highest number of customers will be attracted to,
Implementing demand forecasting that works
The first task for demand forecasting is deciding which data sets are worth including. We always counsel a minimization approach where each data set which you add needs to be justified in its inclusion. However, because predicting future demand is very difficult, innovative use of data which give a clearer picture of what next spring's demand looks like can have great value. The important thing is to have a process in place where the inclusion of any data set for the forecasting projects needs to be justified based on its business value.
With your datasets chosen the next step is to prepare the data. It is almost certain that most of the data sets you have will have multiple uses, of which demand forecasting is only one. As long as you have a strict policy about preparing data as it enters your data lake or data warehouse or platforms, this will ensure that all your demand forecasting data have been adequately validated, cleansed and integrated and have the quality that means they are always up to the new task or the model. You can either get your data engineering team to build a pipeline and get your data science consulting team to prepare their data or use an off-the-shelf product (like the Dativa Pipeline API), to do so.
As with any predictive analytics, asking the right questions of the data will make a big difference to the final results. This questioning begins by documenting a definition of the problem, i.e., the specific requirements of forecasting consumer demand in any particular case. The range of choices open to the decision-makers, such as whether to invest more in live baseball or to instead buy a promising drama, need to be documented, as specific questions that require precise, defined answers.
The specific needs of the business should be the basis of the questions you are asking. Is there enough consumer demand to justify bidding for live motor racing coverage? Should we be offering more hours of reality dating shows? These are the kinds of questions you might be asking of the data. The trick is to extract the relevant data you have about consumer demand for these particular genres by asking specific questions and modelling how they are answered from a technical viewpoint.
Drawing a bigger picture
Some companies will have data going back years, and this can be very useful in drawing a picture of customer demand. For instance, were there any factors which predicted the rise of the reality TV genre? Or its child, the dating reality TV genre? Given the importance of genre for classifying the kinds of shows people will watch, to be able to predict new popular genres can give production companies an edge. Of course, the reality TV genre began in 1997 in Holland with Big Brother and then took the US (and the UK) by storm, but only three years later. Let's imagine some new kind of drama sub-genre is unknowingly about to be launched this fall due to the runaway success of a new drama show which is as big a smash hit as Breaking Bad was in its heyday. As one of your competitors will produce the show, your current demand forecasting cannot factor anything about the new show. But in six months time, when you do the next round of demand forecasting, if your data and their predictive analytics are top quality, and include the runaway success of this new show, your demand forecasting should indicate that the possibilities for this new type of genre are enormous and well worth investing in, perhaps before your other rivals do.
Once you have set in motion a data science approach to demand forecasting it is then important to see how well it worked. So now would be a good time to see how the demand forecast for Q2 this year, created last fall, fitted with the "reality" of Q2, with said "reality" being datasets about actual consumer behavior, in so far as you can ascertain this. Comparing forecasts with real information may then assist in getting better initial data sets and asking better questions in the next round of demand forecasting.