Integration
Forecasting in n8n.
A community node that turns rows in a workflow into a forecast with a confidence band. Point it at the column holding your numbers, say how far ahead to look, and write the result back wherever it came from. No Python, no model to host, no training step.
The free tier is 300 forecasts a month on the real TimesFM 2.5 weights, no card.Install
On self-hosted n8n, install it from Settings → Community nodes, or from the CLI:Then add a VecTime API credential and paste a key from your dashboard. Testing the credential calls GET /v1/models, which is authenticated but unmetered — it costs no quota.bash
n8n Cloud only installs verified community nodes. Until this one is verified there, use the HTTP Request node against https://api.vectime.cloud/v1/forecast — the payload is in the API reference.
It reads both shapes a workflow carries
A time series arrives either as one item per observation or as an array already assembled. The node takes either, and in both cases every series in the run travels to the model in a single batched call.ROWS INTO ONE SERIES
What Sheets, Postgres and paginated APIs produce. Set Value Field to units and Group Field to sku, and each SKU becomes its own series.json
ONE ARRAY PER ITEM
What an HTTP Request or a Code node usually hands over. Set Series Field to series; every item becomes one series.json
What comes back
One item per series by default, or one item per forecast step with Split Points Into Items — the shape that appends straight into a sheet. Whichever you pick, point is the median and the band between q10 and q90 is the forecast's own statement of how sure it is. Safety stock, alert thresholds and capacity buffers belong on the band, not on the median.PER SERIES
json
PER FORECAST STEP
json
Options worth knowing
| Option | Default | What it is for |
|---|---|---|
| Date Field | — | Sorts each series by date before forecasting, and labels every forecast step with the date it lands on. Frequency is inferred from the median gap, or set it by hand. |
| Keep Last N Points | all | Trims each series to its most recent N observations, so a long history stays inside your plan’s context window. |
| Max Series per Request | auto | How many series ride in one call. Left at 0 the node reads your plan’s own ceiling from the API — an unmetered lookup — and batches to it. |
| Split Points Into Items | off | One item per forecast step with the quantiles as flat fields — the shape to append straight into a sheet or a table. |
| Simplify | on | Turn it off to also receive quota usage, request ID and latency on every item. |
| Cold Start Retries | 3 | Retries while the model loads. Only a 503 is retried — never a call that was already billed. |
Quota, batching and failures
Quota counts calls, not series: twelve SKUs batched into one request cost one of your monthly forecasts. The node batches by default, so Max Series per Request is what decides the price of a large run — and it keeps you clear of the per-minute rate limit at the same time. A call that fails is not billed, and a cold model answers a retryable 503 that the node handles itself. Turn on n8n's Continue On Fail and a refusal arrives as an item instead of stopping the workflow.Start with the example workflowThe package ships one: 24 months of history for three SKUs, six months forecast for all three in a single call, one row out per SKU per month. Import it, attach a credential, run it — three of your 300 free forecasts.