A PUBLIC RESEARCH NOTEBOOK

Forecasts
meet reality.

Every hour an AI model predicts where Bitcoin, Ethereum and Solana will trade 24 and 72 hours later. We lock each prediction, wait for the outcome and publish the score — the misses too.

A research experiment built on TimesFM, Google’s forecasting model, with news evaluation alongside. Not investment advice.

Explore the source on GitHub ↗ · Code, tests and the research method.

LOCKED RIGHT NOW

What the model has committed to.

The newest 24-hour forecasts, frozen in the public journal before their deadline. Each one is scored when that deadline passes, whatever happens.

The newest locked forecasts are listed in the forecast journal.

3assets under observation
BTC · ETH · SOL
24 / 72hforecast horizons
Issued before evaluation
365daily historical origins
Per asset and horizon

THE QUESTION

Can context make
a forecast more useful?

A convincing story is easy. A useful forecast has to survive comparison, data problems and the passage of time.

01 / OBSERVE

Start with the series.

TimesFM reads 512 completed hourly closing prices. Its forecasts are stored with the model revision, inputs and issue time.

Implemented
02 / UNDERSTAND

Put news in context.

FinBERT and Laya classify headlines. We preserve when each item became available. Laya annotations do not change a price forecast.

Collecting evidence
03 / CHALLENGE

Let reality answer.

Compare forecasts with the actual outcome and simple baselines. A separate news experiment can progress when enough prospective data exists.

Benefit still unproven

THE EVIDENCE / 001

A baseline worth beating.

Historical · exploratory

The first question is deliberately simple: can the model forecast better than keeping the last observed price?

365 paired daily origins per asset
Bitcoin · 24 hoursLower error is better ↓
Bitcoin, 24-hour forecast errorTimesFM error 1.712. Last-price baseline error 1.627. Lower is better.TimesFM1.712Last price1.627

Mean absolute log-return error × 100. This measures forecast error, not investment returns. A common scale is used across all assets and horizons.

MODEL VS BASELINE+5.2%

higher forecast error

Observed coverage of the nominal 80% interval77.3%Raw model quantiles; calibration is not assumed.
Inspect all six results
Exploratory backtest recorded on 17 September 2026. Error is mean absolute log-return error × 100.
AssetHorizonTimesFMLast priceBand coverage
BTC24h1.7121.62777.3%
ETH24h2.4402.30077.0%
SOL24h2.8412.60074.8%
BTC72h3.2052.98372.3%
ETH72h4.5254.22773.7%
SOL72h4.9684.62774.5%
Frozen snapshot · measured 17 Sep 2026 · published 18 Sep 2026Download the data ↓

THE EVIDENCE / 002–008

Then we asked
the follow-up questions.

Shorter horizons, the uncertainty bands, slower signals, and then the news, the state of the market and other coins. Fixed rules, every variant reported, including the ones that went nowhere.

Four more places we looked, 005–008: headlines, market signals and other coins. None of them says which way the price goes. Read what we found

BEHIND THE SCENES

A small system.
A visible paper trail.

Models run on a dedicated research server. The public site delivers a curated, dated snapshot. Browsing this page does not contact the models or the private dashboard.

Walk through the architecture
01Market data & newsSource + observation time
02Forecasts & annotationsPinned inputs + model identity
03Journal & evaluationStored prediction → later outcome
04This public notebookSelected aggregate results only

THE LIVE PAPER EXPERIMENT

$10,000.
Entirely virtual.

The virtual portfolios record decisions, simulated trades and changing balances, with costs, data checks and a reason for every action. The newest one follows the one rule that survived our tests: a slow trend filter, against a rebalanced benchmark.

Follow the portfolios
VIRTUAL · OBSERVING LIVE

Starting capital

$10,000USD

Recorded decisions.
Transparent simulated results.

BUILT TO BE QUESTIONED

The useful part is what we learn.

Follow the decisions, the checks and the things that did not work as hoped.

Read the development journal