BEHIND THE SCENES / THE METHOD
Every prediction
needs a paper trail.
Knowing what a model said is only the beginning. We also need to know what it saw, when it saw it and how the prediction held up.
01 / THE ARCHITECTURE
Two places. Distinct responsibilities.
The research system runs on a dedicated Hetzner server. Python coordinates collection, model inference and evaluation; FastAPI serves the private dashboard. SQLite stores the journal. Docker keeps the web service and a single worker separate, sharing one persistent data volume.
This public website is a static Astro build delivered by Cloudflare Workers Static Assets. A small read-only API also serves validated paper-portfolio snapshots and a forecast journal stored in Cloudflare D1. The research server pushes this bounded export using a separate publication key. The website cannot read private holdings, the research database, the dashboard or model APIs.
- 01
Collect completed observations
Binance spot candles supply hourly BTC/USDT, ETH/USDT and SOL/USDT prices. CoinDesk and Ethereum Blog feeds supply news titles. Source times, observation times and revisions remain distinguishable.
- 02
Issue forecasts; classify headlines
TimesFM 2.5 reads 512 log closing prices and produces 24- and 72-hour forecasts. FinBERT classifies headline tone. Laya, an open-source model running on the research server, adds structured asset relevance, event relevance, tone and event type. It replaced the paid JEV evaluator.
- 03
Keep the original record
Each forecast preserves its inputs, issue time, horizon, model revision and experiment ID. Previously issued forecasts cannot be updated or deleted through normal database operations. A missing model result stays missing.
- 04
Wait for the outcome
Only an observed close at the actual target hour can score a forecast. Comparisons use the same forecast origins. Historical results and prospectively issued forecasts are reported separately.
- 05
Publish a bounded snapshot
The site includes a small, reviewed results file with an explicit measurement date. The paper page additionally reads regularly published virtual balances, trades and decision trails. Visitors cannot place orders or contact AI services through this site.
02 / INFORMATION BOUNDARIES
The future must stay in the future.
A headline published yesterday but first collected today was not available to yesterday’s forecast. We preserve both times, along with the classification time and headline version. Missing news coverage is represented as missing, rather than neutral sentiment.
Prospective news fusion is a separate experiment. It requires at least 120 mature, covered samples over at least 21 days, plus sufficient distinct news, before fitting a model. Training and calibration are chronological, with overlapping target windows excluded at the boundary.
Laya annotations appear alongside FinBERT in the private research dashboard. They do not change a price forecast. The existing fusion experiment uses its original FinBERT and rule-based features; adding Laya requires a new experiment version.
03 / CONFIDENCE AND DATA QUALITY
A certain model can still have bad inputs.
We keep three questions separate: are the observations usable, how uncertain is the forecast, and how confident is the headline classification? Combining them into one impressive-looking score would hide useful information.
Closed candles, consecutive history, timestamp handling, stale-price rejection, input snapshots and point-in-time news eligibility.
Independent venue comparisons, executable spreads and depth checks, explicit USD/USDT conversion, and per-asset restrictions when data checks fail.
Calibration of Laya on independently labelled crypto headlines, calibrated forecast intervals and a measured benefit from news.
A Laya probability is attached to a specific question about the supplied headline. It is not automatically a probability that the headline is true or that a coin will rise. The model’s Q10–Q90 range is a nominal 80% interval; actual coverage must be measured.
04 / EVALUATION
Compare the same opportunities.
The first historical study uses 365 daily forecast origins for each of three assets and two horizons: 2,190 TimesFM forecasts, with persistence and momentum comparisons. The persistence baseline predicts that the last price will remain unchanged.
The primary measure is mean absolute error of the forecast log return, multiplied by 100. Lower is better. Interval coverage is reported separately. This is a forecasting benchmark, not a trading return.
Paired, seven-day-block bootstrap estimates account for some temporal dependence. They still depend on the stability of the period and the choice of block length. The original study was extended from 90 days to a year, so it is exploratory rather than an untouched confirmatory test.
TimesFM had higher point-forecast error than persistence for all six reported asset/horizon combinations. This is evidence about this experiment; it does not establish a future trading advantage.
Explore all six results or inspect the published JSON.
- Experiment
- v1-log512-h24-72-hourly
- Weights
- google/timesfm-2.5-200m-pytorch
- Pinned revision
- 1d952420fba87f3c6dee4f240de0f1a0fbc790e3
- Price unit
- USDT · Binance spot · completed hourly closes
- Measurement date
- 17 September 2026
LIVE LEARNING
Freeze, settle, check, then learn.
The forecast journal shows each original target beside its observed outcome, direction, error and interval coverage. Scores and source revisions are preserved. Independent Coinbase and Kraken closes, converted with same-hour USDT/USD, check whether an outcome is eligible for calibration learning.
A separately versioned shadow candidate fits a small correction to past TimesFM errors after at least 120 qualified examples spanning 21 days. It uses a chronological training/calibration split and excludes overlapping training targets. Its first future forecasts begin a fixed 28-day comparison against TimesFM, persistence and momentum. Missing observations and weak evidence remain visible; no candidate is promoted into trading automatically.
The existing FinBERT news correction is tracked separately. Laya annotates headlines and does not change numeric price forecasts. The paper-buy veto of the first two portfolio runs reads only the JEV annotations recorded before the switch.
05 / LIMITATIONS
What this proof of concept cannot tell us yet.
- A working model, server and dashboard do not prove predictive value.
- One venue can contain stale or distorted prices. Current checks do not detect every such problem.
- Headline tone is not the same as market impact. Historical news is excluded from the price backtest.
- Overlapping forecast horizons are dependent. Hundreds of forecasts are not hundreds of independent market regimes.
- The historical period begins close to the model’s release; training-data overlap cannot be ruled out from these results.
- Paper portfolios now include explicit fees, spreads, delays and book-depth constraints. They still omit queue priority, full market impact and the behaviour of real exchange accounts; simulated profitability would not prove real trading profitability.
Method documentation reflects the recorded project state on 19 September 2026. This public page is not a real-time service-health report.
PRIMARY REFERENCES
Google Research · TimesFM ↗Pinned model family · model card ↗ProsusAI · FinBERT ↗Vercel · structured evaluation ↗Binance · market data ↗