Displacing Excel Monkeys: How Facebook is Automating the Forecasting Process
When people think about global supply chains, Facebook probably isn’t the first company that comes to mind. Facebook, however, operates a complex global supply chain, delivering its products to billions of people and collecting trillions of data points every day. Companies like Facebook were born in an age of explosive technological change and, in many respects, have inherently digital supply chains. However, as the pace of innovation and competition increase, these companies must continue leveraging digitalization to improve their supply chains at every step.
Supply Chain and Forecasting at a Consumer Internet Company
The consumer internet supply chain follows a familiar model: teams plan for new products and features, forecast demand, develop these products and ultimately distribute them across their user bases. Forecasting sits squarely in the center of any consumer internet company’s supply chain, and is critical for three primary reasons:
- Forecasting allows companies to predict demand, which in turn helps to predict server, storage and system capacity needs. Accurately forecasting these variables helps to reduce costs, avoid downtime, and deliver the optimal product experience to users.
- Forecasting provides an “organic baseline”, which can be used to compare against actual results and assess the efficacy of products over time.
- A strong forecast can prevent teams from spending days and weeks investigating fluctuations in demand that could be explained by seasonality or holidays. These investigations often cause bottlenecks in the development supply chain and can waste millions of dollars of resources.
Why Digitalization Matters in Forecasting
Forecasting is typically a costly and highly manual process whereby entire teams of “forecasting experts” utilize historical data to predict future demand. However, this process can be subjective and inefficient, and an opportunity exists for companies to embrace digitalization in this crucial step of their supply chains.
What is Facebook doing about it? Meet Prophet
In February 2017, Facebook introduced “Prophet”, a forecasting tool available in Python and R designed to “make it easier for experts and non-experts to make high quality forecasts that keep up with demand” [1]. Prophet utilizes an additive regression model that analyzes historical data to identify seasonal trends, shifting holidays, and other trend anomalies [1]. What this means in practice is that teams can feed historical data into Prophet and receive a highly accurate organic forecast. Teams can then manually adjust these forecasts for growth limitations or other irregular factors as they see fit.
Compared to the highly manual process of creating granular forecasts “by hand”, Prophet provides companies with a significantly more accurate and efficient process and removes the need for entire teams of “forecasting experts”. This ultimately allows the entire supply chain to function more efficiently, as the company can better predict demand and control costs, and product development teams can more effectively assess the impact of new releases and optimize their efforts going forward.
While Google and Twitter have launched efforts in their own right [1], neither product approaches the problem as directly. Google’s CausalImpact, focuses more on “estimating the causal effect of a designed intervention on a time series” [2]. Twitter’s AnomolyDetection follows a similar path, focusing on “anomalies in system metrics” [3] rather than providing a high-quality forecast.
What should Facebook do next?
Prophet is not without limitations. The following opportunities could take Prophet from semi-automated forecasting tool to world class supply chain software:
- Reduce the need for human input: Prophet currently requires users to provide access to databases and input key historical events such as shifting holidays and product launches. Facebook should enable Prophet to automatically pull back-end data and evaluate that data more effectively in order to account for these atypical events. Prophet relies too heavily on the forecast owner’s institutional knowledge, which makes it challenging to deploy at scale.
- Expand down the supply chain: Prophet could be used to inform product development decisions in addition to operational performance. Using predictive analytics, Prophet could identify engagement trends before they happen and inform the product development supply chain.
- Increase accuracy and granularity: As Prophet gets “smarter”, Facebook can utilize machine learning to help create more accurate forecasts with even more granularity. Imagine a world where Prophet could forecast engagement down to the county, provide estimates on data center usage by location, and suggest the most efficient methods by which to deliver data across the network. This could drive cost savings down the supply chain and improve site speed and product experience by providing data more efficiently to users.
Outstanding questions:
- How could a tool like Prophet be applied in industries outside of technology?
- As we rely more on machine learning and AI for back end business functions, how do we maintain control and transparency in the way our companies operate?
(Word count: 792)
BIBLIOGRAPHY:
[1] Sean J.Taylor and Ben Letham, “Prophet: forecasting at scale,” Facebook Research (blog), February 23, 2017, https://research.fb.com/prophet-forecasting-at-scale/, accessed November 2017.
[2] Kay H. Brodersen and Alain Hauser, “CausalImpact,” Github, Publishing Date Unknown, https://google.github.io/CausalImpact/, accessed November 2017.
[3] Author Uknown, “AnomalyDetection,” Github, Publishing Date Unknown, https://github.com/twitter/AnomalyDetection, accessed November 2017.
Very interesting piece. Two ideas for how they can develop this are in digital advertising and in fashion and entertainment. First, it could help companies assess the effectiveness of their digital advertising. Understanding the true value of digital advertising is hard because its often not clear what a real sales baseline would be without the ads and because its hard to attribute growth in sales to digital adverts (while it’s easy to get data on clickthrough rates – it’s very difficult to get data on actual purchases). Facebook’s Prophet tool and its other advertising data would help companies create a reasonable baseline and then test adverts against it more accurately. Second, Facebook could apply Prophet to the fashion and entertainment industries. Facebook contains a trove of user information on current trends. Its Prophet tool could help companies understand and forecast developments in these trends and popular products and stay ahead of the curve.
Excellent essay and analysis. In addition to the points you’ve made about Prophet’s power on the demand forecasting front, I wonder how this type of software might be applied on the purchasing side of organizations to drive procurement efficiencies and buying power through better understanding of how externalities work together to affect market prices. Said another way, could companies leverage the speed and power of this software to analyze historical price fluctuations correlated with a multitude of market events to determine the optimal time to purchase inputs for their supply chains, as well as the optimal quantities?
Thanks Michael for a fascinating post and yet another example of how human tasks are being increasingly done by digital technology. Prophet seems to be an excellent product – however, I believe there are limitations to the forecasting ability of big data based regressions. Firstly, you need a large amount of historical and representative data to be able to draw good correlations and subsequent forecasts, and this may only exist in a few situations. Secondly, a model, that relies on correlation but does not understand causality, may not be able to account for fundamental changes that could mean the future forecast is very different from what the past would have predicted.
It’s interesting that Google’s CausalImpact seems to have taken a different approach, and maybe a combination of both may be the answer. But for good or bad, there still seems to be some time to go before excel monkeys with the ability to understand causality can be entirely done away with.
This commentary was insightful and timely, as an article about CFOs moving their teams away from Excel came out just this past week in The Wall St. Journal and makes for good reading [1]. I agree with Kieron that fashion may prove a useful application for Prophet or other smarter forecasting tools, but encountered lots of potential issues when looking at Anaplan, Adaptive Insights, and some other Excel replacements when at my last job. The crux of the issue is that once your core model is built, you still need to adjust for lots of different variables. These adjustments can take just as much time in these other programs as with Excel, and just as Michael discusses with technology, in retail you have product / store launches, promotional periods, holidays, weather changes (weekend snowstorms in Q1’15 impacted our bookings about ~5%), etc. Any model is only as good as the inputs, and my concern with Prophet and other products like it is that people assume the “machine learning” going on captures everything and therefore don’t pay enough attention to the inputs. Eventually we will certainly reach a place where a holistic algorithmic model understands your business and the linkages to endogenous / exogenous events better than any human-built model. But I don’t think we are there yet, and the hurdle to learn a new program vs. keep plugging away in Excel hasn’t been crossed (for most applications).
1. https://www.wsj.com/articles/stop-using-excel-finance-chiefs-tell-staffs-1511346601
Thank you for your thoughtful commentary and analysis, Michael. This is certainly a fascinating case of digitalization in the supply chain that could have a profound impact on a number of industries. In addition to the limitations you mentioned, I see another challenge is that you are essentially developing your forecast through a ‘black box’. As more companies turn to this to replace the traditional finance function, auditors will have an even more difficult time reviewing these already subjective analyses. If Prophet and tools like it can be used to replace the human element of forecasting, how do we impose safe guards that will prohibit manipulation and implement appropriate oversight procedures? [1]
[1] Adam Compain, “AI, machine learning key to increasing forecasting accuracy,” https://www.joc.com/international-logistics/logistics-technology/ai-machine-learning-key-increasing-forecasting-accuracy_20170319.html, March 19, 2017.
Great post – as an excel monkey myself can’t help but be concerned I may be out of a job in several years. To what extent has backtesting been done on automated forecasting technology? Would be interested to backtest ML forecasting relative to the human equivalent. Also, I’m interested to see how this changes (devalues?) the role of the private equity and investment banking professional. If a machine can make educated projections, why pay analysts? Both businesses are forecasting related at the entry level but relationship based at the top so interested to see how this technology changes the hierarchy of these organizations – will the bottom of the pyramid (analysts and associates) even be necessary in the future?