I recently took an online course from Wharton on Operations Analytics and it was valuable because it was for traditional businesses, not game companies. Social game companies are great at player, marketing and customer analytics, to the point where it is hard to eke out competitive advantage using analytics because everyone is so good; it is largely the cost of doing business. To gain an edge it is sometimes useful to look outside the space and learn best practices from other industries, even ones considered less sophisticated than online gaming, and apply these insights to our business.
While Operations Analytics is a focus for retailers and bricks-and-mortar businesses, it often does not make it on the radar for social, mobile and iGaming companies. The foundational problem in the space is even referred to as the Newsvendor Problem, clearly based on a very traditional business. After taking the Operations Analytics course, I realized many useful applications for gaming companies from this field.
Forecasting demand and allocating resources
I have worked at many large gaming companies, including several that were public companies, and, despite the companies reliance on analytics, financial forecasting was at best a guess, at worst a hope. A retailer, especially a land-based one, lives or dies by its forecasts. Order too much and you are left with potentially worthless inventory. Order too little, and you miss profits that can then be used to market, grow or sustain during seasonal periods.
Accurate forecasting is critical to make valid decisions today (hiring, investing, M&A, etc.), and it is as critical for online gaming companies as it is for a bricks-and-mortar business. For example, a good forecast will help you hire the right size Customer Support team. It will also help you optimize cash flow, manage your marketing expense, allow you to seek financing preemptively, determine when to launch new content and features and more.
The first key learning from Operations Analytics related to forecasting is that point forecasting are usually wrong. Point forecasts are unreliable not only in business but in all elements of life. The chance of predicting the amount of rainfall next month accurately is close to zero.
Instead, a good forecast should be a range, showing the likely outcomes. When modeling an uncertain future, you build probability distributions based on past data (and possibly incorporating expert estimates). You can put likelihoods on different scenarios, such as low, normal and high. When looking at your potential sales for an online game (or your portfolio of games), which is almost a continuous distribution of scenarios (like amount of rainfall), you want to group the scenarios rather than create individual ones, a continuous distribution.
One of the core ways of using existing data is to base forecasts on moving averages, a practice I recently implemented and have seen quite useful. A moving average is a forecast based on the average of the n most recent observations. In my case, we have created much more accurate predictions looking at the past 28 days of data (so n=28), taking both the mean and standard deviation. It is not a perfect tool, though, as it misses trends (i.e. pandemics) and does not show causation.
While finding the mean and standard deviation of past data is a good foundation for creating your forecast, sophisticated companies integrate additional sources of information. These may include an aggregation of forecasts (for example from each of your product managers), customer surveys (how much they expect to change their purchasing behavior), a jury of executive opinion (what your leadership team thinks) or the Delphi Method (individual opinions that are compiled and reconsidered again and again until there is a consensus).
Smoothing data for seasonality
Another common practice in operations analytics that is not always used in online gaming is smoothing for seasonality. While game companies are keenly aware that sales (and other KPIs) will be impacted by day of the week or hour of the day (referred to as seasonality), some companies are better than others at adjusting for this seasonality.
Most game companies realize these differences and will not compare Saturday revenue with Sunday revenue. Instead, they will look WoW (week on week) and compare Saturday with the previous Saturday. The weakness in this approach is that a lot happens in seven days so you are losing very useful information in the comparison, rather than seeing immediate trends you are dealing with a seven day delay.
An approach to avoid this problem is to remove seasonality from the numbers. You can do this by looking historically at the numbers, creating a sample mean, then finding seasonal averages (if we are looking at day of the week you would have a mean for Mondays, a mean for Tuesdays, etc), divide the seasonal averages by the sample mean to create a seasonal factor and then divide each observation by the seasonal factor. By using a seasonal factor, you can identify immediately any big deviations and not have to wait a week to understand if your game is broken.
Projecting new game launches and other high uncertainty situations
While online game companies are not always sophisticated in modeling their forecasts, it gets much worse when projecting for new games. Some companies rely on the wishful thinking strategy, projecting what revenue will be if the game is a hit (despite the fact less than 20 percent of new games are successful). Others take the opposite tack, assuming no revenue from new projects until they already have data. Neither of these approaches provides useful information for companies’ planning. They either muddy the water so that people do not take the financials seriously or provide misleading data that leads to too much or too little additional resources (such as hiring support staff or preserving marketing budget).
While there is significant variance in potential outcomes for new games, there are methods to create a usable range for forecasting purposes. Given that there is limited (possibly a soft or beta launch) or no actual demand data, you need to augment that data with other sources:
- Sales of similar games. How have comparable products done at launch and over time.
- Composites. Estimates from the product team or marketing.
- Customer surveys. Responses from target customers to market research.
- Jury of executive opinion. Estimates from your executive or leadership team.
- Delphi method. Individual opinions compiled and reconsidered, repeat until overall group consensus is (hopefully) reached.
You then should refine your estimate based on past experience. Look at previous estimates and compare them to actuals. If you are historically 5X overly optimistic or 20 percent overly pessimistic traditionally, adjust your estimate to reflect this historical variance. Also use the historical data to calculate a standard deviation, so you can create a forecasted range of potential values, rather than a specific number. Then incorporate this range in your planning.
A more advanced method to use when estimating a new product launch, or anything else where there is a high level of uncertainty, is simulation. You can build a simulation of a new product launch using a third party tool (StatPlus is one such tool) coupled with the projections above. You can then see 100 or 1,000 or more likely outcomes with a distribution similar to past product launches. This simulation will then not only give you a range of likely outcomes but also help you identify best and worst case scenarios, so you can accurately gauge risk and reward. You can then optimize your internal resource allocation (i.e. hiring new support agents or performance marketers) by optimizing for reward while using your risk measures as a constraint.
For me, the most valuable insight from operations analytics is that you should only have one goal. When creating an optimization model, you optimize for one decision variable (revenue, profit, cost, etc.), changing multiple variable within constraints. It ends up being a, sometimes complex, algebraic model with a solution. Mathematics such as this, however, can also have a broader application.
I will never forget an executive committee meeting I had several years ago where a new CEO said we have two goals, grow top line revenue like a growth company and grow margin. At the time, several people asked which is more important and the CEO responded they both are, we need to optimize for both. Fast forward several years later and the company wallowed in mediocrity, without significant organic revenue or margin growth.
So what does that story have to do with the math behind optimization models? In an optimization model you can only optimize for one objective. You can optimize for profitability or revenue but if you try to optimize for both, Excel would give you a big error message saying it is impossible to optimize for both. Instead you can optimize for one and set the other as a constraint (optimize for growth with margin >= 20%). This mathematical reality is also a strategic reality, when setting your objective variable do not set multiple objectives but make the hard decision as to what is most important and then use the others as constraints when building your strategy.
- While Operational Analytics are a focus primarily in retail and traditional businesses, there are many best practices that iGaming and social game companies can leverage.
- Forecasting is central to generating and earmarking resources but is often a challenge for game companies, rather than trying to create a point forecast create a range based on moving averages and looking at standard deviation. For new products, create a simulation that will show the distribution of potential outcomes and the risks and rewards possible.
- You need one, and only one, distinct goal and then optimize your strategy around that goal; it’s impossible to optimize for multiple goals. Use constraints to incorporate what used to be additional goals.