The Five Koan of Beverage Forecasting
Having a hard time figuring out how to use forecasting when the results never seem good enough? For many of us, forecasting seems like either too much work, or too complicated to get right. So we fall back on our best estimates, based on our reps’ information, or simple spreadsheets that use some historical sales data to generate trend lines.
The answer may lie in unexpected places. Zen Buddhists have a name for it -“Koan”. Koan refers to a paradoxical riddle that demonstrates the limits of conventional logic and provokes enlightenment. With deep apologies to philosophers and history, here are five Koan that may help you in your quest to improve your forecasting:
1. Simplify your models.
The beverage industry is complex and changing fast. SKUs are exploding as suppliers look for new growth amid changing consumer preferences. Distributors are pressed to order, sell, service and rotate SKUs effectively in this environment. Amid this change, a good sales forecasting model won’t last. If you’ve built a great model with lots of inputs and conditions, the fact is that it will be harder to adapt to new market conditions. It’s safer to keep your forecasting models to about 10 to 12 inputs and monitor the data quality for those inputs over time. You may give a little on predictive power, but you’ll gain a better long-term view of your business and have an easier time explaining your forecasts to management.
2. More data is a bad idea.
In the age of big data, you may be advised to throw all your data into the big analytical melting pot and let the computer sort it out. This can work for trained data scientists who know when the ‘recipe’ is working and when it’s not. But for the rest of us, we should really understand what is driving our forecasts and seek the most relevant data for our purpose. In the beverage industry, there are five key data categories that should be driving your models:
- Historical shipments
- Sales rep input
- Pricing, promotion, and local competition
- Distributor depletions data
- Retail point-of-sales
While there are only five categories, together they represent a lot of data. This is where having a forecasting system is essential. Spreadsheets are not that great for modeling the data, doing what-if analysis, or storing and tracking historical results.
3. ‘Forecastability’, not accuracy.
Everyone focuses on accuracy, generally setting some arbitrary benchmark, like “I want my models to be 95% accurate.” The fact is, your accuracy is only as good as the predictability in your data. Forecasting on the performance of mature products is easy; forecasting new products with new packaging is extremely hard. An analyst or forecasting software can provide you the key metrics to know how “good” your forecasts can be – given your data. Remember, forecasts are just a decision support tool. Even if accuracy is relatively low, the directional information you get may help you avoid big mistakes or find unique opportunities.
4. Time series analysis is not forecasting.
Forecasting is not an algorithm in a piece of software. Forecasting is what you do with the information you gather to make the best sales, production and inventory decisions. It’s a process, and the statistical forecasts are one (important) input. Step back and look at how you are running your forecasting and how your management makes decisions. A good process has five steps:
- Pull fresh data
- Generate baseline model
- Integrate input from field reps
- Collaborate, revise and review model
- Publish model for use in planning
5. Don’t let the machine do all the work.
In a busy world, it’s tempting to want a piece of software that can take on the whole process and output a perfect forecast. Software has come a long way to make the process easier and allow you to input more diverse kinds of data to see what improves accuracy. But if forecasting were perfect, we’d all be rich. Take your forecast and integrate it with your best judgment. Keep your people involved in the process to ensure buy-in, real-time feedback, and insights that are not revealed by the data in your model.