The informal consensus problem
Here's how most restaurant groups make major forecasting decisions. The owner, the GM, and maybe a senior buyer get in a room. Someone says "I think this summer is going to be strong based on how last April felt." Someone else says "I heard the neighborhood is getting a new hotel two blocks over, that should help." A third person says "the economy seems soft, I'm not sure we should stock up."
They make a decision. Sometimes it's right. Sometimes it's wrong. Nobody tracks the calibration over time, so nobody gets better at predicting.
This is not a critique of the people in that room. They're using the information they have. The problem is that the information they have is systematically biased in a few predictable ways.
How experience anchors on the wrong things
Human memory of past events is not a flat distribution. We remember outliers. The record-breaking Saturday in October 2022. The brutal two-week stretch in January when nobody came in. The week that Taylor Swift played MSG and business was inexplicably dead because everyone was at the show.
When we use "how last April felt" as our anchor for planning this April, we're using a recollection that is heavily weighted toward those memorable outliers, not toward the median week in April.
The median week is what you should plan around, with a distribution of possible scenarios around it. The memorable weeks are useful data for scenario planning, not for baseline forecasting. Most operators do this backwards.
What structured forecasting actually fixes
When I talk about forecasting tools for hospitality, I'm not primarily talking about sophisticated data models. I'm talking about replacing informal consensus with structured processes that force specificity and accountability.
The simplest version: before you make a major inventory or staffing decision, write down your prediction with a confidence level. "I expect our July revenue to be between $180,000 and $210,000, with 70% confidence." Then track whether you were right.
Doing this for 12 months gives you calibration data about your own forecasting accuracy. Most operators discover they're overconfident (their 70% confidence intervals contain the right answer less than 70% of the time) and that they systematically overestimate summer and underestimate fall.
Knowing your biases is more valuable than having a better model. You can correct for known biases. You can't correct for biases you don't know you have.
Prediction markets as a forecasting layer
Prediction markets are a specific tool for aggregating distributed forecasts. Rather than one person's gut or a room of people's informal consensus, you're aggregating the probability-weighted predictions of many people with different information.
For industry-level questions, this is genuinely useful. What will mezcal's on-premise share be at the end of 2026? Which regions will see the strongest restaurant opening activity? Will tip credit legislation pass in the five states currently considering it?
These are questions where many people have partial information. A bartender in Austin knows something about mezcal's momentum in their market. A Chicago restaurateur knows something about the local real estate environment for new openings. A New York labor attorney knows something about the legislative prospects for tip credit changes. A prediction market aggregates all of that distributed knowledge into a single probability estimate that's more calibrated than any individual expert view.
The Liquor Bets premise
Liquor Bets is built on the thesis that the spirits industry would make better decisions with better forecasts, and that better forecasts are available through structured aggregation of distributed knowledge.
The specific questions in the market are designed to be: specific enough to have a clear resolution, relevant enough to matter for business decisions, and early enough in their development that there's genuine uncertainty and genuine information value in the forecast.
The market doesn't tell you what to decide. It tells you what the crowd of informed participants thinks is most likely to happen. You combine that signal with your own knowledge and your own risk tolerance to make the actual decision.
The value isn't the prediction. It's the calibration. Knowing that the crowd puts 65% probability on the X outcome, rather than just hearing "everyone in the industry thinks X," gives you a much more actionable signal about how much conviction to put behind that view.
Where I'm taking this
The long-term vision is a forecasting layer for hospitality that operators can actually use, not just an academic market. Short-term sales predictions for neighborhood events. Staff planning models calibrated to local data. Category trend signals for buyers and sommeliers making allocation decisions.
The infrastructure for this exists. The data is mostly available. The gap is in making it accessible and useful for operators who are too busy running their businesses to become data scientists on the side.
That's the problem I'm working on.
