# How it works > How each forecast is built from the company's own filings and calls, how the release grades it line by line, and what the words on each page mean. Canonical: https://pricedin.markets/how-it-works/ ## How it works Our goal is a detailed understanding of how each forward estimate is built: the numbers, the assumptions under them and the facts each assumption rests on. That gives investors a clear frame to judge how far their own assumptions are from what is priced in. ### Before the release For each company, a committee of seven AI analysts forecasts the income statement line by line for the next two quarters, and sales and adjusted earnings per share for this year, next year and the year after, each with a low and a high case. Each page shows the forecast's cutoff, to the second. A new forecast is a new page; the earlier one stays. ### Where the numbers come from Seven AI analysts read what the company itself says: every 10-K, 10-Q and 8-K, every earnings call and every company presentation, for the last five years. Today they are Opus 5.5, GPT-6.1 Sol, GLM 5.3, DeepSeek V4.1 Flash, Kimi K3.1 Pro, Gemini 4 Argon and Grok 4.7. Each forecasts alone, from the same filings, the same tools and the same cutoff, and a second model checks each forecast for errors before it is final. Then each reads the other six forecasts once and may change a number, but only by naming the line of a filing that changed its mind; code finds that line again before the change counts. A page names the analysts by letter, A to G, never by model. They also see the economy as it stood at the cutoff: 160 public economic series from the St. Louis Fed's database (FRED), among them oil and fuel prices, interest rates, inflation, wages, business activity, housing, exchange rates and markets. Each value is the one published before the cutoff, so a later revision never leaks into a forecast. The list is the same for every company, and no analyst is told which series matter. The company's history under the forecast comes from its filings, and each of its numbers is checked against the filing. The headline is **one analyst's whole forecast: the middle analyst**, the one nearest the middle of the seven on both sales and earnings per share. So the headline, the base case and the line table are one forecast and always add up. The ranges are the middle of the seven analysts' values at each point. The **80% range** runs from 1 in 10 to 9 in 10, so the reported number should land inside it about 8 times in 10: the [track record](https://pricedin.markets/track-record/) counts how often it does. Inside it, the **50% range** has a 1 in 4 chance of a result below it and 1 in 4 above. Where one analyst sits far from the others, the page shows that view and its reasons. ### We do not guide the analysts We do not guide the analysts in any way. Every analyst gets the same instructions for every company, with no example from any company. Nothing in them says what to conclude, how much weight to give a fact, or which facts matter. The items selected to explain each forecast (its assumptions, the facts under them and the signs to watch) are decided entirely by the analysts. Each names its own drivers and the facts they rest on. The writer picks among them and adds none. Code checks each assumption against the words of the analyst it comes from, and finds each fact's quote again in its source. ### The words on the page A writer model writes each page's words from the seven forecasts. It writes no number of its own: code checks every number in its words against the forecast's figures, or against a quote from the filing or the call, and finds every quote again in its source. ### Minutes after the release An AI reads the release when it reaches the SEC. Code keeps a number only when it is in its quote from the release, when each reconciliation from a reported figure to an adjusted one adds up, and when last year's numbers agree with last year's 10-Q. Then the result page goes up: each line against our forecast, a beat, a miss or in line by its effect on profit. When the 10-Q comes, its tagged facts grade each line again. We never narrate the stock's reaction. ### After the call When the call ends, the seven forecast again from the new quarter, the same way. The page after the call shows each number that moved, beside the forecast from before the release, with the reason and its quote. ### Adjusted figures Adjusted earnings per share follow **the company's own definition**, as its release reconciles it. A dash is never an estimate. ### Our number, not the Street's Not resold Wall Street data. Our number is our own committee's: seven AI analysts who read the company's own words and public economic data; we give them nothing from Wall Street. No access advantage, no waiting for a revision cycle. On this site "consensus" means the Street's estimate only, which the [track record](https://pricedin.markets/track-record/) sets beside ours after the release. ### The record From 9 October 2026, every forecast and every result stays up, hits and misses alike. Each result adds a graded row, and the track record page will show them all.