Skip to content
World Finance Terminalby GodEngine
Open terminal

GodEngine

Not one model. An engine with two ways of answering.

TimesFM 2.5 produces the quantile forecast around a single series. The GodEngine swarm resolver answers the other half of the question — which futures are on the table, and how heavy each one is.

RTSCurve forecastSCHEMATIC · NO DATA
FREDUS TreasuryGodEngine

TimesFM 2.5

temporal/ · wft/forecast/

A foundation time-series model, vendored byte-verbatim from god-engine and verified against a lockfile. Any series the product holds can be sent to it, with horizon and transform chosen in the lab.

Answers how wide is the distribution around this series

Feeds Forecast Lab · forecast columns · Divergence · Record

Swarm resolve

wft/godengine/client.py

GodEngine’s resolver streams branching futures with probability mass, assumptions and triggers per branch. It is the other half of the engine, not a different product.

Answers which futures are on the table and how heavy each one is

Feeds Scenarios · Simulation · Regimes

What happens when you ask for a forecast

Six steps, and what stops each one.

  1. Series

    A named series is selected — from yfinance, FRED, CoinGecko or any bound connector.

    Refuses when the connector is unbound

  2. History

    Observed history is read from Postgres, with its as-of stamp from the last successful fetch.

    Refuses when history is shorter than the context window

  3. Weights

    The forecast worker loads the model file already verified on local disk.

    Refuses when size or sha256 does not match

  4. Forecast

    TimesFM 2.5 returns a median and the quantile band for the chosen horizon.

    Refuses when the decoder is the polyfit fallback

  5. Record

    The forecast is written down with horizon, quantiles, baseline and decoder before any outcome exists.

    Refuses when never — recording precedes display

  6. Render

    The panel draws observed in data ink and the forecast in engine ink, with provenance beside it.

    Refuses when decoder_name is StubDecoder

What every forecast carries

A plausible line is worse than no line.

The decoder name travels to the browser with every forecast, so a run from the fallback decoder is visible rather than hidden — and it draws nothing at all.

decoder_name

Which decoder produced this. The fallback is visible, not hidden.

horizon

How far ahead, fixed at write time and never re-chosen after the fact.

q20 / q50 / q80

The quantile band drawn as the fan.

baseline

Random walk, scored alongside so skill is measurable.

as_of

The server's last successful fetch of the input series, never the browser clock.

source

The named connector behind the input series.

run_id

Ties this forecast to its row in the ledger when it resolves.

Weight integrity

Vendored, not forked.

01

temporal/ is a byte-verbatim copy of god-engine’s TimesFM layer.

verified against temporal/VENDOR.lock
02

A modified vendored file fails the build rather than shipping quietly.

scripts/vendor_sync.py --check in CI
03

Weights are fetched from a local path or a release, never from git.

scripts/fetch_weights.py
04

Size and sha256 are checked against the recorded checksums.

weights/CHECKSUMS.txt
05

Only a verified file is moved into place, atomically.

925,181,104 bytes

Six commitments the engine makes

What the engine will never do.

Quantiles, not a point

A single line implies a confidence the model does not have. Every forecast ships q20, q50 and q80.

A baseline travels with it

Skill is measured against a random walk. Without a baseline a forecast can only be admired, not judged.

Written before scored

The horizon is fixed at write time. Nothing is re-chosen once the outcome is known.

The fallback is loud

A polyfit stub is plausible and therefore dangerous. It draws nothing and says why.

Vendored, not forked

The model layer stays read-only and verified; product behaviour lives beside it, not inside it.

Engine ink only

Forecasts never borrow the colour of an observed number, in any theme or colourblind mode.

Where the engine appears

FCST

Forecast Lab

Pick a series, horizon and transform; read the fan and its provenance.

REC

Track Record

Resolved forecasts, coverage, calibration and skill by horizon.

DIV

Divergence

Where the engine most disagrees with the baseline, and whether that paid.

SCN

Scenarios

Branch tree and probability mass from the swarm resolver.

SIM

Simulation

Push a shock and watch it propagate through the linkage graph.

MKT

Every tape section

An engine column beside the observed one, never in the same ink.