Mycroft: From Reading the Past to Forecasting the Future

From transcript extraction to forward guidance forecasting

[0:00]Five episodes in, Essis could read an earnings call and grade itself on what it read. This week it stopped waiting for the call. I'm Anjana. Here's how a system that only ever looked backwards started looking forward. Essis is a system that reads earnings call

Mining features from decision logs without new APIs

[0:15]Transcripts and extracts financial guidance signals automatically. It runs four independent readers on every chunk of every transcript, keyword matching, FinBERT sentiment analysis, named entity recognition, and a large language model.

Pre-registering append-only predictions before earnings calls

[0:30]A triangulator fuses their outputs. Every signal is pre-registered before anyone knows the outcome, then graded against the market at 30, 90, and 180 days. The system does not trust any single reader or any single confidence number. It triangulates, scores, and self-corrects. That is the foundation.

Two-level scorecards: Grading predictions vs CEO statements and market response

[0:50]Here is what comes next. Until now, Essis read the past. Transcripts came in, signals went out, but the signal history itself contains patterns. What direction did company A report last quarter? How many consecutive quarters has it raised? Has the stock already moved before the call? What are peer companies saying this quarter? These are features, and they are all computed from

The forecast dashboard and testing your own decision history with Claude

[1:11]Data the system already has. No new sources, no new APIs, just the decision log it has been building all along. A logistic regression model takes these features and predicts where guidance will go next quarter before the CEO opens their mouth. The prediction follows the same honesty rules as every extraction. It is logged before the earnings call happens. Ticker, predicted direction, confidence, the full feature vector snapshot, model version, and timestamp. Locked. Append only. After the call, the transcript is extracted and the actual guidance direction is known. Only then does the scorecard grade the prediction. Did the model get it right? The prediction cannot be revised, softened, or deleted after the fact. The same discipline that governs extraction now governs forecasting. Now there are two score cards asking two different questions. Level one, did the prediction model correctly forecast what the company would say? Predicted raised, actual raised, correct. Predicted maintained, actual lowered, wrong. This grades the prediction against the extraction. Level two, did the extracted guidance correctly predict

Stock performance? Raised guidance, stock outperformed, correct. This grades the extraction against the market. Two levels. The prediction the guidance. The guidance predicts the stock. Each level is graded independently. A prediction can be right about what the CEO said and still wrong about what the market did. The dashboard now has a forecast view. For every company with an upcoming earnings call, you see the predicted direction, the confidence, which features drove the prediction, and whether it agrees or disagrees with what the rest of the sector is saying. A timeline shows how accurate past predictions were for that company. One screen to see what the system expects before the call happens. From extraction to prediction. The system that read the past now forecasts the future. Essis, episode six. Let's recap with Claude. The features come entirely from the decision log the system was already keeping. No new sources. No new APIs. A logistic regression model turns them into a prediction and that prediction is locked before the call. Exactly the way

Every extraction always was. Then two score cards grade two different things. Whether the model predicted the guidance and whether the guidance predicted the stock. The system that read the past is now forecasting the future and it's grading itself on both. Your turn. I keep a record of decisions I've made and how they turned out. A log, a spreadsheet, a journal, anything with a history in it. Can you help me? One, name three patterns in that history that might predict my next decision before I make it using only what's already in the log. Two, tell me which of those patterns I could actually check against a real outcome later, and which are just vibes. And three, ask me honestly whether I'd be willing to write down my prediction before the outcome, so I can't quietly revise it afterward. Paste that into Claude and find out whether your own history is predictable, and whether you'd let it be graded. From reading the past to forecasting the future, that's in Jona.