Forecast Anything

Local forecasting

Forecast with your assistant.

Connect a CSV and run Timestral locally through your assistant. Compare the forecast with a seasonal baseline and see what changes as new data arrives.

Local Timestral · CSV input · MCP

Daily demand

Synthetic demo

Daily orders at the forecast cutoff8 observed daily values, 1, 3, 1, 3, 1, 3, 1, 3 orders, through 23 Oct 2026 UTC. For 24 Oct 2026, Timestral predicts 1.013 orders and the seasonal baseline predicts 3.000 orders.Orders0123423 Oct · cutoff16 Oct24 Oct 2026
Timestral
1.013
Seasonal baseline
3.000

Next target

provisional

Photo by Nenad Stojković · CC BY 2.0 · Cropped, monochrome

01Same-target comparison

The next daily demand point

Synthetic CSV 296 observations

Daily orders at the forecast cutoff8 observed daily values, 1, 3, 1, 3, 1, 3, 1, 3 orders, through 23 Oct 2026 UTC. For 24 Oct 2026, Timestral predicts 1.013 orders and the seasonal baseline predicts 3.000 orders.Orders0123423 Oct 2026 · cutoff16 Oct20 Oct24 OctDaily orders at the forecast cutoff8 observed daily values, 1, 3, 1, 3, 1, 3, 1, 3 orders, through 23 Oct 2026 UTC. For 24 Oct 2026, Timestral predicts 1.013 orders and the seasonal baseline predicts 3.000 orders.Orders0123423 Oct · cutoff16 Oct24 Oct
ObservedTimestralSeasonal baseline

02Saved question

When new data arrives.

The worker picks up the new CSV row and issues another forecast for the saved question, even after the assistant disconnects.

The target moves from to . The two forecasts cover different days and aren’t directly comparable.

Read the recorded demo guide
Worker update
New observation7 orders ·
New target
Statusprovisional

The update is provisional: there’s only 1 new point to check against. The watch requires 30 new evaluation origins before it qualifies again.

Backtest, provenance and record IDs

Mean absolute error in orders, measured at matched historical origins.

PartitionOriginsTimestralBaseline
exploration300.2692.000
selection300.0082.000
final300.0092.000

The fixture is synthetic. These scores do not establish real-world accuracy.

Timestral 28M ran locally on CPU. The seasonal-naive baseline is an independent comparator. The worker processed the update while the assistant was disconnected.

Question
d5b8f6fcb6c14d7ef478177febdc44d7c3c6445c54378819bd2a365c1a0b20bb
Original run
ff66ccf8e568170951c6f24e2eb9382d12aa2a1cafceb88ac20c1e8a122d06cf
Successor run
c9126611b3a3925829999ec0a49e02548a80bed8128e38d05540a1c3240590a1
Original snapshot
5be928f372298567b1844c48cc2e36528bd09543b71264782b047fb52d71131a
Updated snapshot
b9645a3b8c75c50ee0f7783b4bf4d8d11ff3c71096234a0b63211535a142646a
Backtest
0a6bfb9ee5613c1f7704dae2d057b18d0c342eae91b91ede3247588fe3e4870b
Weights revision
49b26409ac87b5935539a5d8ffe7f397b720db47
Code revision
a534bdbd244d64f3d9d743f46c7bf6fc330ed4ff
Fixture SHA-256
8c13c1e16ac46e1ec9b62785035ba52d3aa4f51268593d5e37616000434bab27
Provisional reason
insufficient_new_final_origins

03Local setup

Connect your assistant

  1. Install the locked runtime and MCP dependencies from a reviewed checkout.
  2. Review the Timestral weights terms. Run setup to prepare the model, start the local service and register it with Codex.
  3. Enable the registered MCP server in Codex and ask it to call get_diagnostics.
Full installation guide

From the repository root

Requires macOS, Python 3.12+, uv, the Codex CLI on PATH, and a checkout kept at a stable path.

uv sync --locked --extra mcp
.venv/bin/forecast-anything setup \
  --workspace ~/forecast-workspace \
  --prepare-model
codex mcp list
Preview the optional console
Forecast Anything console showing a dependency graph and a recorded Timestral forecast.
Timestral forecast on synthetic data in the local console.Open full screenshot Console setup guide