← All articles · September 2026 · 3 min read
How to Make Better Predictions: 7 Evidence-Based Habits
Base rates, decomposition, calibration and honest scoring — the methods forecasting research keeps pointing back to.
Methods Beat Hunches
A recurring finding in forecasting research is that how you think matters at least as much as how much you know. In the Good Judgment Project, a research team led by Philip Tetlock and Barbara Mellers that competed in a forecasting tournament run by the US intelligence research agency IARPA from 2011, the strongest forecasters were not defined by credentials. They were distinguished by habits: starting from base rates, using precise probabilities, updating often, and keeping score. Tetlock and Dan Gardner describe them in the book Superforecasting (2015). Here are seven habits you can adopt.
1. Start With the Base Rate
Daniel Kahneman and Amos Tversky called it taking the "outside view". Before you think about what is special about this case, ask what usually happens in cases like it. Planning a software launch? Look at how often similar launches in your company shipped on time. The planner Bent Flyvbjerg built "reference class forecasting" on this idea after finding that large infrastructure projects routinely overrun their budgets. Start from the base rate, then adjust for the specifics — not the other way round.
2. Break Big Questions Into Parts
A question like "Will we launch in Europe this year?" is hard to estimate in one go. Split it into steps that all have to happen: regulatory approval, a local partner, a translated product. Estimate each, taking into account how they depend on each other. If all three are required, the chance of the whole cannot be higher than the chance of the weakest step, and is usually lower than any of them. Decomposition often reveals that an intuitive "80%" was too optimistic.
3. Use Numbers, Not Words
"Likely" means 55% to one person and 90% to another. Write a probability. Superforecasters tend to use fine-grained probabilities (65% rather than "probably"), and the precision forces you to think about how sure you really are. It also makes your predictions scoreable.
4. Update in Small Steps
A forecast is not a promise to defend. When new information arrives, adjust — usually by a little, occasionally by a lot. Tetlock's research found that the best forecasters revised their forecasts more often than others, typically in small increments. Write down why you changed your number; the reasons are as useful to review as the number itself.
5. Look for Evidence Against Yourself
Once we hold a view we notice what supports it. Before you commit, ask: what would I expect to see if I were wrong, and am I seeing it? Argue the other side for five minutes, or ask a colleague to. This is the most uncomfortable habit and one of the most useful.
6. Keep Score Properly
Track every prediction with its date, probability, deadline and outcome, including the embarrassing ones. After a few dozen, check calibration: of the things you called 70% likely, about 70% should have happened. A Brier score (the squared difference between your probability and the 0-or-1 outcome, averaged) gives you one number to watch over time. A spreadsheet is enough.
7. Make the Record Hard to Edit
A track record is only convincing if it could not have been cleaned up afterwards — by you or by anyone else. For predictions you may want to point to later, fix the wording and the date somewhere you cannot change. icalledit.io does exactly that part: it writes a SHA-256 hash of your written prediction into its own Bitcoin transaction and gives you a certificate and proof.json that anyone can check without us. It does not score or resolve predictions; your spreadsheet or forecasting platform does that.
Putting It Together
For your next important forecast: define the outcome and the deadline, find the base rate, break the question into parts, write a number, note what would change your mind, and log it where it cannot be edited. Review it on the deadline. Repeat. Improvement comes from the feedback loop, not from any single clever insight.
A timestamp proves that content existed at a point in time. It is not legal advice, a patent, a trademark or a copyright registration. Compare the options.