Performance Forecaster

Forecasts that learn. Every prediction compounds.

Felix predicts revenue 90 days ahead, retrained nightly on your accounts. Once enough past forecasts are checked against what came in, each range is resized to match how often it held.

How Felix works
Leads withThe confidence band, and what would move it.
NeverGive a point estimate alone, or quote an accuracy it has not scored against actuals.
Why this exists

Forecasting that breaks every Monday. And never gets sharper.

Models go stale instantly.

An analyst builds a forecast on Monday. By Wednesday, the inputs have shifted. By Friday, nobody trusts the numbers. Repeat weekly.

Cycles take weeks, not minutes.

Pulling data from five platforms, cleaning it, running regressions, sanity-checking. 15 to 20 hours per forecast cycle that could be 15 seconds.

Misses do not improve next time.

Last quarter's forecast missed. Did the model learn? No. Same model, same blind spots, same miss next quarter. Just with newer data.

Trends arrive late, not early.

Seasonal shifts, competitor moves, creative fatigue. A spreadsheet cannot detect emerging trends three weeks before they hit revenue.

What Felix ships

Five capabilities. One forecasting workflow.

Each capability maps to a specific job: predict the line, check it against what came in, follow the trend, flag what moves off its usual level, and hand the forecast to the rest of the agents.

90-day revenue forecasts

Revenue, ad spend and order forecasts, each with a range that is recalibrated against Felix's own track record.

  • Month-end and quarter-end forecasts from the same fit
  • Confidence ranges, not single numbers
  • Updated nightly on your accounts

Example. A range for next month's revenue, wider the further out it reaches

Compound learning

Every forecast is checked against what actually came in, and the record sets how wide future ranges are.

  • Backtest history kept per account
  • Each checked forecast records its error and whether the actual fell inside the range
  • After ten checks at a horizon, its ranges are calibrated on that record

Example. Every miss is compared against actuals and used to calibrate the range

Early trend detection

Refits the trend in your own history every night, so a change in direction shows in the forecast as soon as your data carries it.

  • Weekly patterns and changes in trend
  • Revenue, ad spend and orders, each fitted on its own
  • Fitted on up to 180 days of your own history

Example. A slowdown in orders shows in the next morning's forecast

Anomaly alerts

Felix checks yesterday's Meta spend against its usual level each morning, and new findings reach the Daily briefing.

  • Meta spend that moves well off its usual level
  • A month forecast below its plan appears as a finding in the Daily briefing
  • Anomalies are matched against your active goals

Example. A Meta spend spike flagged in the next morning's briefing

Agent-aware

Reads the synced store and ad data Dana checks, and Scenarios and Goals read Felix's forecast as their baseline.

  • Scenarios start from Felix's current forecast
  • Goals read the forecast to say whether a target is in reach
  • Only days from connected, healthy sources are used

Example. A scenario on the Scenarios page starts from Felix's forecast range

A typical week

What Felix does while the team sleeps.

One workweek, hour by hour. Specific numbers vary by account.

Mon · 4:32 AM
Trained

Runs nightly retrain on the last 180 days of account data. One Holt-Winters fit each for revenue, ad spend and orders.

Mon · 5:14 AM
Forecast

Forecasts revenue 90 days out. Each range narrows as past forecasts are checked against what came in.

Mon · 7:30 AM
Flagged

Flags when the month's revenue forecast falls below plan.

Mon · 9:18 AM
Anomaly

Flags yesterday's Meta spend as well off its usual level and matches it to active goals.

Mon · 11:02 AM
Read

A scenario on the Scenarios page starts from Felix's forecast range.

Tue · 5:10 AM
Reconciled

Checks every forecast that came due against what came in and records the error.

Next Mon · 10:00 AM
Delivered

The Weekly receipt lists the forecasts that came due, with their ranges and actuals.

Compound accuracy

Cold on day one. Calibrated to your account over time.

Forecast accuracy is not a static number, and it is not one we publish: no figure here has been measured on a customer account. What Felix does is check every prediction against what actually happened and calibrate its ranges on that record.

Week 1

Cold start on your historical data, once 14 days have synced: one Holt-Winters fit with a weekly season per metric, refit nightly.

Week 4

Forecasts that came due are checked against actuals, and each check records its error.

Week 8

After ten checks at a horizon, the ranges at that horizon are calibrated on the record, wider or narrower.

Week 24+

Every horizon calibrated on your own record, and Goals and Scenarios reading the same forecast.

Holt-Winters with a weekly season, refit nightly

What's underneath

Real infrastructure. Not a wrapper.

Felix runs on a purpose-built forecasting stack: one fit per metric, nightly retraining, scored backtests, and a range calibrated on its own record.

Reads from

Meta, Google, TikTok, and Shopify via OAuth. Only days from connected, healthy sources are used.

Meta AdsGoogle AdsTikTok AdsShopify

OAuth-based, no warehouse copy

Reports to

The Daily briefing and the Weekly receipt, in Slack and email.

Slack

Anomalies arrive in the Daily briefing each morning

Collaborates with

Scenarios and Goals read Felix's forecasts. Dana's source health decides which days are used.

Shared memory, not message-passing

Built on

One Holt-Winters fit per metric, simulated 2,000 times for the range.

Nightly retraining, scored backtests every cycle

Ready when you are

See Felix forecast your account, or talk to us about your stack.

The pilot connects your accounts to every agent. Felix's first forecast arrives with the next nightly run once 14 days of history have synced.

Looking for something specific? Parker validates lift →, Dana keeps inputs clean → or Sam tests scenarios →.