How To Forecast Without Historical Data

My Aunt Leslie is a fantastic cook, but she doesn't often cook in someone else's kitchen.

We had Thanksgiving together one year in a vacation rental at the beach. Lots of family were there, and we were all looking forward to what we knew was going to be a delicious meal. Everyone sat in the living room, chatting and catching up, while Leslie put finishing touches on the fixings and reset the kitchen timer for the turkey in the oven.

Popping her head into the living room, she announced, "Hey all - turkey's got 30 more minutes, everything else is ready. Dinner's at 5:30."

"Awesome, I can't wait," I replied.

30 minutes later the timer rings, Leslie pulls open the oven door. "Are you freaking serious?" we hear from the kitchen.

The turkey's raw, oven is off. Chef Leslie pressed the wrong buttons and needs 3 more hours. 8pm it is.

5pm turned to 8pm... the bird is cooking, getting close to done. "30 more minutes," she says, "I have to reheat the potatoes, but I swear this isn't going any later." We're borderline hangry, but dinner has already become a funny story that we lovingly won't let Leslie forget.

8:30pm, the potatoes are reheated, the turkey is done, and we finally sit down to eat. We feast, and the food is delicious as always.

Bayesian Forecasting

Bayesian forecasting is actually a lot like that Thanksgiving dinner, and it can be a great way to forecast when the only data you have is your own judgement and what you see happening in the moment.

The idea is to start with a prior belief about what you think will happen, then update that belief as you get more information.

At 5:30pm, Leslie expected to see a fully cooked turkey. It turns out, the state of the turkey was an outlier because at 5:30pm the turkey was raw. When she checked again at 8pm, the turkey was close to done, but it still took a little longer than she expected. After underestimating twice, Leslie finally calibrated her forecast and we all sat down to eat "on time."

Bayesian forecasting works in a similar way. If you are trying to forecast sales of a brand new product, then you might start with a prior belief that sales will be normally distributed around some mean. You don't know exactly what that mean is, but based on your domain expertise let's say you guess that the mean is normally distributed around $1.5M. This implies your sales forecast for next quarter is $1.5M.

If the product ends up generating $2.5M, then that's evidence that your prior belief about the mean might have been a little low.

What's nice about the Bayesian forecasting math is that it will update your belief about the mean by just the right amount to account for the new information. In this case, your new belief will be that the mean is a little higher than $1.5M. Therefore your next forecast will be a little higher than $1.5M as well.

This approach allows you to make forecasts even when you don't have historical data to rely on. Unlike simpler methods like taking a moving average, it sequentially refines your expert guess and is more robust to outliers. This can be particularly helpful when you are forecasting in a new market or for a new product.

Conclusion

At first glance, not having historical data for forecasting can make it seem like the only way to move forward is to make a wild guess. And that's kind of true! But with Bayesian forecasting, that guess isn't so wild, and it can be a great way make a forecast when you don't have historical data to work with.