Most important decisions get made before all the facts are in. A team signs a supply contract without knowing next quarter's prices. A company hires for demand it can't see yet. Waiting for certainty is a choice too, and it isn't free.
We're building Prolepsia for those decisions. It helps you lay out your options and find the questions whose answers would change your mind, then puts a probability on each one. You record a choice, check it against your goals and hard limits, and set a date or a condition for looking at it again. As the evidence changes, Prolepsia keeps the forecasts up to date.
Prolepsia Research is where we show the work.
What we find in the research
The headline number rarely settles a forecasting question. Here's a recent example. We were forecasting the U.S. crude oil spot price at Cushing, Oklahoma, for 31 December. National crude inventories were falling, which usually points to a tighter market. But Cushing is where the main U.S. crude futures contract is delivered, and stocks there were going the other way. They rose from 21.5 million barrels on 11 September to 27.6 million on 16 October, about 28% in five weeks. A falling national number is weak evidence of a shortage when the delivery point itself is filling up.
That was part of why our forecast's middle value came out at about $100 a barrel, a little under the spot price of about $104 at the time. The other part was the futures market, where contracts for February delivery settled near $98. We put a 4-in-5 chance on the price landing between $75 and $137. The full forecast is public if you want to check our reasoning.
We'll write up more findings like this: the data that mattered, the data that looked like it should matter and didn't, and the places where our forecasters disagreed with each other.
The shape of a forecast
The oil forecast above is usually quoted as one number, about $100. The forecast itself is 200 small probability buckets, each 45 cents wide, between $60 and $150, with a little probability left over at each end. The middle value is $100.50. There's a 10% chance of a price below $75 and a 10% chance above $137.
Those two tails are not the same size. A fall of $30 from the middle, to below $70, has a 6% chance. A rise of $30, to above $130, has 13.5%. The chance of ending above the entire range, past $150, is 5.8%. The chance of ending below it, under $60, is 1.8%. The risk in this forecast sits on the upside, and no single number can tell you that.
The distribution answers a buyer's question directly. If your budget breaks at $120 a barrel, the chance of that happening is 21%. Prolepsia reads that chance off every forecast of a question tied to a hard limit, and it can flag your decision for review when the chance rises past a level you choose.
When the question resolves on 31 December, the forecast is scored with a rule that charges for distance. A price of $100 would cost about 0.06. That isn't zero, because the forecast spread its probability out, and spreading has a price. A price of $90 or $110 would cost about 0.08. A price of $130 would cost about 0.20, and $70 about 0.23. A miss costs more the farther it lands, and the cost grows faster than the distance. The rule is public, so you can run the same numbers on any of our forecasts.
From a probability to a choice
A probability means little on its own. Say there's a 20% chance a key supplier misses a deadline. That might be easy to live with, or it might be the biggest risk you face. It depends on what a missed deadline costs, what a backup supplier costs, and whether you can wait a month and know more.
We'll work through cases like this, where costs and hard limits do the deciding. Sometimes the option that looks best on average breaks a hard limit, and then the average doesn't matter. Often a single piece of evidence would settle the choice, and the useful work is figuring out which one.
The decision stays yours. Prolepsia keeps the reasoning in one place and helps you see when it needs another look.
Keeping score in public
As outcomes come in, we'll go back to earlier forecasts and compare them with what happened, misses included. When we get something wrong, we'll say how wrong, and whether the reasoning or the evidence was at fault. We'll link our sources and state our assumptions, so you can argue with the analysis instead of taking it on trust.
Our first public questions have already resolved, and the record is mixed: two right, two wrong. In Brazil's presidential runoff, we gave Flávio Bolsonaro a 67% chance of winning, while the betting market Polymarket gave him about 80%. Lula won with 51.8% of the vote. We were wrong, but less wrong than the market, and the reason was in our report. Datafolha's own polls had moved toward Lula three times in a row, from 52–48 to 51–49 to about 50.5–49.5. We read that as a reason to doubt a clear Bolsonaro win, though not enough to call it for Lula.
The track record now lists every resolved public forecast with its outcome and score.
Subscribe on Substack to follow along. If you have a forecasting question or a decision you'd like us to look at, write to us at contact@prolepsia.com.
Your Prolepsia Team