How Probable works

Most news tells you what just happened, then leaves you to guess what it means or what comes next. Probable does the opposite: we focus on what’s likely to happen, using actual data instead of pundits’ opinions.

We don’t tell you what to think. We tell you what the evidence says is most likely, and we show our work so you can decide for yourself.

The number you see at the top of every Probable story — “Probable’s read: 72%” — is our own forecast, not anyone else’s. We arrive at it by weighing four classes of live evidence against the historical base rate for that kind of question. None is sufficient on its own; the synthesis is the editorial product.

Prediction markets. When real people put real money on whether something will happen, the resulting prices are remarkably accurate forecasts — often more accurate than polls, pundits, or expert panels alone. We pull live odds from several widely-used prediction markets — most often Polymarket, alongside Kalshi (CFTC-regulated and US-legal) and Manifold. We cite these markets as a data source only; we do not direct readers to any of them as a place to trade. Markets are usually our most-cited input because they integrate everyone else’s views into one price.

Professional analysts. What Wall Street’s research desks, central bank projections, and expert panels are saying. That includes the Federal Reserve’s own Summary of Economic Projections, bank research from firms like Goldman Sachs and JPMorgan, economist surveys from Bloomberg and Reuters, political analysts at the Cook Political Report, and aggregated superforecasters at the Good Judgment Project. These voices carry weight when they line up with markets and even more weight when they don’t.

Public opinion polls. Polls measure things markets can’t — what voters say, what households expect, what consumers feel. We use poll aggregators (FiveThirtyEight, Cook), well-methodology individual polls, the University of Michigan consumer sentiment survey, the New York Fed’s consumer expectations data, and political forecasting models. Polls move slower than markets but they measure different things.

Official data. When the question involves the economy, government policy, or measurable real-world events, we go to the original source: the Federal Reserve, the Bureau of Labor Statistics, the Treasury, the CFTC, the SEC, the Census Bureau, the IMF, central banks abroad, and the official press releases of the agencies and officials involved. This data is public, and we link to it.

History (base rates). A market price tells you what traders expect today; it doesn’t tell you how often events like this actually happen. Most “will it happen by [date]” questions resolve no, so before we settle on a number we factor in the historical base rate for that kind of outcome. We hold a base rate for each class of story — monetary policy, elections, geopolitics, and so on. To be precise about where those numbers come from today: they are starting estimates, set when we launched, based on how often that kind of question has historically resolved yes. We have built the machinery to recalibrate them against our own resolved forecasts, and we will, but we have not run that recalibration yet — we haven’t been publishing long enough for our own sample to mean much. When our base rates first move because of our own track record, we’ll say so here.

The way it comes together: we identify the day’s most-watched questions, gather what each input says, and produce Probable’s own number — we don’t simply echo a market. We weigh how strong the live evidence is: a deep, heavily-traded market with several sources agreeing is trusted almost entirely, so our number stays close to it; a thin market, or a question no market is pricing, carries less weight, so we lean more on the historical base rate. That’s why you won’t see many 50/50 coin flips — a genuine coin flip only survives when the live evidence is both strong and truly even. We attach a confidence level — high, medium, or low — that’s narrow when the evidence is strong and wider when we’re leaning on history. The published number is Probable’s own reasoned read. We also compute a second number by fixed formula — it blends the live sources for that story toward its historical base rate, mechanically, with no judgment involved — and we keep that formula number beside our own as a cross-check, not as the answer. When the two disagree sharply and the market behind the formula is deeply traded, that story goes to a human before it publishes, and we show you both numbers and explain the gap. The number is always grounded in the sources we cite — never the AI’s unsupported guess.

A word on how we gather those last two classes: prediction markets and official data come straight from their public feeds, while the analyst and poll inputs are drawn from the named analysts, institutions, and pollsters quoted in the reporting we cite, plus sources we enter by hand. We surface and attribute those views; we never lift copyrighted or proprietary data.

We do not republish anyone else’s reporting or research. We summarize, we cite, and we link out to the originating source so readers can audit our reasoning.

A first draft is not what gets published. Before a briefing reaches you it goes through a review layer, and because we ask you to trust our numbers, we think you should know exactly what that layer is and what it isn’t.

An independent editor reads every draft. Once the draft is written, a second AI model — one that had no hand in writing it and no stake in defending it — reads the draft against the same source material and hunts for specific failures: claims that aren’t supported by any source, missing attribution, opinion or betting language creeping in, questions phrased so vaguely they could never be scored, and prose that reads like a machine wrote it. It also writes the strongest honest argument against our own read. A third pass then rewrites the draft to fix what the editor found and to fold that counter-argument into the story, so you see the other side rather than a one-sided case. This step can only change words, never the number.

A red team argues our number is wrong. The number gets the same adversarial treatment. After we land on a probability, a separate model is pointed at it with one job: using only the sources we already cited, make the strongest possible case that our number is too high or too low, and say which way it should move. A final pass weighs our original reasoning against that challenge and settles on the number we publish — moving it only as far as the evidence actually justifies, and rewriting our stated reasoning to match so the words and the number never drift apart. Every story where this process moves the number is flagged and looked at by a person before it publishes.

What the human does, precisely. One person — the same person who runs Probable — reads each briefing before it goes out. That read is a proofread: tone, clarity, whether anything is obviously off, whether a story reads as partisan or as betting advice, plus a closer look at anything the system has flagged. It is not a fact-check. We do not verify every claim, every link, or every probability against its underlying source, and we won’t claim otherwise — it isn’t possible for one person at the volume we publish. Anyone telling you their AI publication has a human verifying every fact is describing something other than what they are doing.

So: several layers of automated adversarial review, and one human reading for tone and obvious problems. That combination catches a great deal. It does not catch everything, which is why the correction log and the scoreboard exist and are public.

Let’s be straightforward about this: Probable is primarily an AI-driven publication. The vast majority of our writing is generated by large language models from Anthropic, working from data and reporting we feed them. We think AI is a powerful tool for cutting through partisan noise and writing clearly about complex topics — but it also means errors are part of the deal, and we’d rather tell you that than hide it.

Here’s how it actually works:

• The AI works only from the data and reporting we provide it, never from its own memory or assumptions. The prompt explicitly forbids it from making any claim that isn’t grounded in a cited source.
• More than one model is involved, and they are set against each other on purpose — one writes, others are tasked with finding what’s wrong with the writing and with the number. That’s described in full above.
• A human proofreads each piece before publication — checking for tone, clarity, and obvious red flags. We do not, and cannot, fact-check every claim, every link, or every probability against its underlying market data. That’s beyond what one person can do at the volume we publish.
• This means errors will sometimes get through. We expect this, and we plan for it. Our public correction log catches what we and our readers find later, and we keep it open and visible.

We think AI in journalism should be transparent, not hidden — which is why we’re upfront about using it instead of pretending otherwise.

We don’t give financial advice. Probable is journalism, not investment recommendations. Even when we discuss markets, we’re describing what’s likely to happen, not what you should do with your money.

We don’t claim to be 100% accurate. Forecasting is inherently uncertain. The markets we cite are usually right but sometimes wrong. The analysts we cite are sometimes wrong. The polls we cite are sometimes wrong. Our own synthesis is sometimes wrong. AI sometimes makes mistakes. When we get something wrong, we’re happy to admit it and fix it publicly on our scoreboard.

We don’t tell you what to think politically. Our approach is probabilistic and evidence-based, not partisan. If you find us leaning one way or the other on a political question, that’s our failure — please tell us so we can correct it.

We don’t promote betting. Prediction markets are our data source, not our recommendation. We are an editorial product, not a betting platform.

When we get something wrong — and we will — we’re happy to admit it and fix it publicly. We make it easy for you to see how often that happens.

Our public scoreboard tracks our actual track record: how many of our forecasts turned out right, how many turned out wrong, and what we missed. It’s automatically generated whenever a question we wrote about resolves (a market closes, an event happens, a number is released). Calls we got right post to the scoreboard automatically, with no human step. Calls we got wrong are held back only long enough for us to write and publish the correction that goes with them, so the miss and the explanation appear together rather than the miss appearing bare; if a correction can’t be produced, the miss is released on its own anyway. A wrong call is never quietly dropped.

Where a question can’t be settled automatically — no market closed it, no clean number landed — it’s flagged and a person resolves it by hand rather than being left open indefinitely.

If you spot a mistake we haven’t caught — a fact that’s wrong, a source we’ve misrepresented, a probability we cited incorrectly — please tell us. Email corrections@theprobablenews.com or reach us through the contact link below. We respond as quickly as we can and post visible corrections on the original piece.

Probable does not have a political viewpoint, and we work hard to keep it that way. We don’t endorse candidates. We don’t editorialize about whether outcomes are good or bad. When we explain why a market is moving the way it is, we describe what traders and reporters say is happening, not what we wish was happening.

The only “side” we’re on is being right about probabilities. If our analysis tells you something you don’t want to hear — about an election, an economic outcome, a geopolitical event — that’s by design. We trust you to handle the data.

Probable does not sell your email or your reading habits. We use industry-standard email and analytics tools (Beehiiv for our newsletter, lightweight privacy-preserving analytics for the site) which keep your data secure. The full privacy policy will live at /privacy.

Probable is an independent media project, launched in 2026. We’re not owned by a larger media company. We’re not funded by any political organization, party, or candidate. Our only revenue comes from reader subscriptions, newsletter sponsorships, and affiliate partnerships, all of which we disclose openly.

If you want to support Probable, the best thing you can do is subscribe to the newsletter — it’s free, and it keeps us going.

• General questions: hello@theprobablenews.com
• Corrections and tips: corrections@theprobablenews.com
• Sponsorship inquiries: sponsor@theprobablenews.com
• Press and media: press@theprobablenews.com

Probable is for informational purposes only. Nothing on this site is intended as investment, legal, tax, or financial advice. Prediction markets carry risk; if you choose to participate in them, you do so at your own risk and following the laws of your jurisdiction.

Probable is primarily an AI-driven publication. Our content is drafted with the assistance of large language models from Anthropic. A human proofreads each piece before publication but does not fact-check every claim against its underlying source. Errors can and do happen. When we catch them — or readers flag them — we correct them publicly on our scoreboard.