The forecast lives in someone's head
"Last month was this, so this month will be about the same" — and next month's budget gets planned on intuition rather than on data.
Forecasts that are later checked against fact
Valkiria forecasts impressions, clicks, conversions, revenue, budget, audience, churn and growth from your own history alone, and stores each one with its confidence and source references. When the real numbers arrive, the forecast is compared with them and scored for accuracy.
Why a forecast usually costs nothing
Nobody checks the AI forecast
The model states a number with confidence, and a month later nobody comes back to see whether it held. The error is never found, and the same model predicts again.
The number has no source
A tidy figure with no note of which data produced it, or what was missing from that data, can be neither challenged nor defended in front of a client.
How a forecast is built and verified
Feed it the history
The module hands the model your accumulated performance. Its instructions forbid going beyond that data and require listing sources, assumptions and any missing evidence.
Generate the forecast
One button builds forecasts for the period — 30 days by default. Each record is stored with its type, category, confidence and an unverified-estimate status.
Check it against fact
When the period closes, the actual values and an accuracy score are attached, so you see which predictions deserve trust and which do not.
What predictive analytics gives you
Eight forecast types
Performance, budget, audience, revenue, CTR, conversions, churn and growth, filtered by category in a single list.
A confidence on every record
A forecast with no confidence is labelled "needs evidence", and how confidence moves over time is drawn on its own chart.
Back-testing built in
Forecast and actual sit side by side, accuracy is scored as a percentage, and the average across all forecasts tells you what the model is worth.
A factuality badge per card
The status shows whether sources back the forecast: an unverified estimate can never be applied automatically.
A forecast you can budget with
Expected impressions, clicks and conversions for the next 30 days come as a chart you can put behind a spend plan.
Your choice of AI provider
The Ollama → z.ai → OpenAI → Anthropic → Qwen chain is configured per install, so nothing is hard-wired to one vendor.
In numbers
- 8
- forecast types, from CTR to churn
- 30
- days — the default forecast horizon
- 0
- forecasts applied without a human
Questions about ad forecasting
What is the forecast based on?
Only on your own historical performance, which is passed to the model with the request. The instructions require it to return predicted values, a confidence score, source references, assumptions and a list of the evidence it was missing.
Can I trust the confidence number?
Confidence is itself an estimate by the model, not a measurement, so it is never the last word. Forecasts are compared with the actual values, and an average accuracy accumulates next to them showing what that confidence is worth in practice.
Can a forecast raise a budget or switch a campaign on by itself?
No. A forecast is stored as an unverified estimate, and that status blocks automatic application. Campaign changes are made by a person in the campaign modules, where anything new is created PAUSED.
Which metrics can be forecast?
Eight types: performance, budget, audience, revenue, CTR, conversions, churn and growth. The list filters by category, so forecasts for one client or one direction can be read on their own.
Does my data go to OpenAI?
That depends on your configuration. The module runs through a provider chain set up per install, which can start with a self-hosted Ollama; nothing is hard-wired to a single vendor.
Build a forecast you can verify
Forecasts are built from your data, labelled as estimates and never applied on their own. You can check the accuracy after the very first cycle.