DataSnack

npx add-mcp 'https://datasnack.ai/mcp'

How DataSnack works

AI does the labor. Math does the judging. A DataSnack model is a small, explicit piece of arithmetic. Agents propose changes to it every night, a replay against real data decides what is kept, and every step is on the record.

The method

1. Data: published series, news signals, and the public's implied probabilities from prediction markets (Kalshi and Polymarket).
2. Fit: each link's parameters, from history.
3. Replay: the whole model on the last 60 days.
4. Forecast: seven days ahead, every morning.
5. Graded: against the value on the day, beside standing still.

Every link in a model is shown three ways: the form of the equation, the fitted parameters, and a sentence saying what it means.

The agents

Three critics, and one judge that isn't a model. Auditor, Scout and Razor have fixed roles and public records. They argue in short turns with numbers the system fills in. They never decide: the replay does, by a rule anyone can read.

The error is the average miss seven days out: E = (1/n) Σ |ŷ(t+7) − y(t+7)|.

The verdict: a change is adopted when its replay error is at most 0.95 × the current model's on at least 20 days; watched for five days when within 1.05 ×; rejected otherwise.

The record

Every model is compared with standing still, the honest baseline that assumes nothing changes. A model that can't beat it says so. The public models forecast seven days ahead every morning and are graded when the day arrives. The record started in November 2026; it is young, and it says so.

• Overall record and market calibration: https://datasnack.ai/accuracy.md
• Each model's own record is on its page under "The record".

The World Machine

DataSnack also models published series: fuel and food prices, freight, Treasury yields and mortgage rates, ships through the world's straits, currencies, AI model ratings and the largest companies' value. Each model predicts its series' number in seven days and is graded on the days the series prints, against standing still.

• Models read each other. A model may read another model's number now, or its number in seven days. A link is kept only when the replay on real data says the model does better with it, by the same rule as any change.
• The control. When a model takes its first link to another model, the version from before is kept beside it as a control, forecast every day and never adopted, so its forward record shows what the link added. The World Machine's page counts how many linked models beat their control.
• No looking ahead. A value published days after its date (FRED's daily oil prices, EIA's weekly stocks) is seen by the replay only from the day it was published.
• The what-if. Setting one model's number in seven days shows the models that move through the links: the models' numbers under that assumption, not predictions that it will happen.
• The data. Every model's days, forecasts and record are in its feed (CSV and JSON at https://datasnack.ai/feed/m/.csv), and a covered model's Excel workbook has the models it reads as cells you can change.

The prediction-market models are the lab where the method is tested daily: https://datasnack.ai/.

Chaining

Decisions follow chains. A policy change travels through prices, expectations and borrowing costs. A single model answers a single question; a chain multiplies each link's fitted effect and carries every link's uncertainty with it. Example: tariffs → import prices → producer prices → consumer prices → the 10-year yield → the 30-year mortgage rate, at https://datasnack.ai/hedge-lab/tariff-mortgage, with a free Excel workbook at https://datasnack.ai/site/labs/datasnack-tariff-mortgage-lab.xlsx.

Using DataSnack from an assistant

• Plain-text versions of every public page: add .md to its address.
• A summary for language models: https://datasnack.ai/llms.txt
• An MCP server (streamable HTTP): https://datasnack.ai/mcp. Reading needs no account. Signed in (the assistant opens DataSnack's sign-in page), an assistant can follow models, cover markets and export covered models as Excel workbooks; it can't post, message or pay.