Jingsong Cui on Forecasting in the Television Industry

In this special Q+A, we talk with Jingsong Cui, clypd’s Head of Media Analytics. We discuss the importance of forecasting, clypd’s approach to forecasting, and what the future brings to forecasting.

What’s your background?

I have been working with data and models throughout my career. In graduate school, I studied economics and used econometric models to analyze social and economic data. After getting my Ph.D,  my first job was working for a marketing research company called @Futures. I led a team of statisticians to build forecasting models for pharmaceutical clients. The company was acquired by Nielsen in 2010. Within Nielsen, I worked with several industry verticals, across Buy (which focuses on consumer spending) and Watch (which focuses on media consumption). I have always enjoyed doing applied research and using data to solve real business problems.

Why did you join clypd?

While I was part of Nielsen’s Data Science group, working with the commercial side of Media Analytics, I led the development of an audience forecasting model for a cable network client. The objective was to improve the accuracy and efficiency of the forecasting model as used by the client. In the end, the model was able to deliver results with 20 – 40% better accuracy (see my article in Nielsen’s Journal of Measurement), which has clearly demonstrated the value of forecasting models for the TV industry.

Even with these convincing results, it was still just a concept. To really tap into the power of an audience forecasting model, we needed to implement the model, with all the mathematical formulas and computing algorithms on a technology platform through an automated process.

Around the same time, I heard that clypd was looking to expand its data science team and invest in analytics and research products, including forecasting.  After learning about the company, I felt very excited. clypd is reshaping the TV advertising landscape by leveraging unique technology. For forecasting, clypd provides the ideal platform to build the next generation forecasting model. I was very fortunate to have the opportunity to join the team.

What makes clypd unique?

First and foremost, the people at clypd. We have some of the brightest industry thought leaders. We also have people who have built successful technology companies with a strong background in digital advertising and software development. Within clypd, we all work closely together.

The second unique advantage is our technology platform. At clypd, we have built a robust platform designed to help media owners and advertisers execute their advertising campaigns efficiently and effectively.

Why is forecasting important?

At clypd, improving forecasting is essential for client success and a useful forecasting model should be designed to work in harmony with the business operation of the TV industry. There are a few important things to keep in mind when talking about the business side of TV advertising.

First, the majority of TV inventory is still sold through the Upfront. The Upfront market works like a futures market where media owners sell their advertising inventory in advance, by several months or even over a year. These transactions usually come with some guarantees. This puts a lot of pressure on the sellers (media owners) to improve their forecasts and minimize the risks of both underestimating and overestimating.

Second, a significant portion of the inventory is sold through scatter planning, where media owners sell their inventory in the nearer term, by days or weeks. Scatter planning requires the forecasting model to be more efficiently implemented, so that results can be available much more quickly. The shorter gap also means more information could be available – e.g., who will be the guests in a talk show, how recent episodes have performed, what have competitors being doing, etc.

It is also worth pointing out that innovations such as connected devices and smart TV are opening up opportunities to customize ads at household and person level. This requires a forecasting model to identify the most valuable and relevant target consumers.

How does clypd approach forecasting?

At clypd, our forecasting model is built with these considerations:

  • Enable advanced targeting, beyond Age and Gender
  • Employ a variety of data sources
  • Incorporate latest techniques
  • Work seamlessly within our technology platform and with other services

As advertisers move away from traditional age and gender-based targeting and towards advanced targeting based on behavioral, attitudinal, and even psychographics data, we have been improving our forecasting models to enable this change. Our forecasting model can use either clypd or a media owner’s forecasts by age and gender, and add another layer to generate advanced target forecasts.

There are many data sources available now to enable advanced targeting and each has its own pros and cons. clypd has been working with the industry through the Advanced Targets Standards group (www.ATSG.tv) to create guidelines around best practices when working with these data sets.

On the methodology front, clypd’s  data science team have been improving our forecasting capabilities. We use machine learning to process a large amount of data to identify the most useful signals and patterns for forecasting purpose. Natural language processing (NLP) is useful for incorporating text-based information (e.g. description, cast and crew) to improve forecasts. Data mining is useful to identify special events that can have a special impact on future audiences and therefore need to be correctly accounted for.

Another unique aspect of our forecasting model is how it interacts with downstream optimization work. We believe a well-designed forecasting model should take into consideration how the results are used, and incorporate the end-user’s objectives in the model development stage. To put it simply, model development should be guided by the use case. In our case, media owners and advertisers use our forecasts to optimize ad campaigns for specific target segments. The success of forecasts here is related to the granularity of the inventory packaging – the programs, selling titles, network and day-part combinations that are the building blocks of the plan.

Lastly, for a well-thought-out model to be truly useful, it also needs to be implemented on a robust platform, and work smoothly in the production environment. Our engineers have built and tested our platform to make sure it scales well for more inventory sources and multiple campaigns running in parallel. With our platform, media owners can automate their campaign planning process and forecasting has become an integrated component of that process.

What does the future of forecasting look like?

Internet of things (IoT) and Artificial Intelligence (AI) seem to be getting all the attention these days. I think they will also have a big impact on how we do forecasting and run business. But it is important to know their limitations and use them wisely.

In a nutshell, forecasting is about trying to predict what is going to happen in the future. But looking at the past may not always be useful to inform about the future. Remember the  Google Flu Trends debacle?

For an industry as creative as TV, it means we need to combine technology with human wisdom. Economists like to use the phrase “animal spirits” to describe the uncertainty of human emotions and behaviors. While it makes forecasting work more challenging, it also makes it more interesting. I look forward to finding out how TV forecasting will evolve in the next few years – I am sure clypd will play an important role in that.

Leave a Reply