NewWeather Demand Modelling is live. Forecast demand before it arrivesWeather Demand Modelling is live
AI Insights

Intelligence that models your business, not just charts it.

Forecasting, anomaly detection, churn and demand modelling, segmentation and reconciled attribution, in one engine. It learns how your business actually trades, then tells you what is happening, why, and what to do.

Models fit per account · reconciled to source · honest when data is missing

Generated insightLive
Your Connoisseur Core segment is drifting toward lapsed, with roughly $140k of annual margin at stake over the next quarter.
Why. Repeat cadence in this cohort has slowed past its at-risk threshold, and the slowdown leads the revenue dip by about three weeks, so it is visible before it lands in the P&L.
RecommendedBuild a win-back audience for this segment and measure the recovered margin.
What the intelligence does

More than dashboards. Models.

Each capability runs on your own data, reconciled to source. These are the models working under the surface of every Blufire edition.

Forecasting

Demand and revenue, modelled forward.

nowP90P10

Twelve months of revenue and true margin with confidence bands, not a single hopeful line. Event multipliers for launches and sales.

Method. Holt-Winters smoothing, P10 / P50 / P90 bands from rolling residuals.
Weather demand

Which weather moves your demand, and when.

0d1d2d3d4d5d6d7dr=.78

The platform discovers which conditions drive your demand and how many days ahead, per location, without a rule written for your industry.

Method. Deseasonalised Pearson correlation at 0 to 7 day lags.
Anomaly detection

Spikes and drops, against your own volatility.

flagged

Surfaces the changes that matter relative to your normal range, with a likely cause attached, and stays quiet on the noise.

Method. Bollinger bands and z-score, actionable only.
Segmentation & personas

Behaviour-true segments, not demographic guesses.

Customers scored and clustered into named, interpretable personas and margin-true value tiers, from how they actually buy.

Method. RFM percentile scoring and k-means clustering.
Churn & lifecycle

Who is drifting, on your real cadence.

Every customer placed in active, at-risk, lapsed or dormant on thresholds learned from your own purchase rhythm, not a generic 90 days.

Method. Per-account recency thresholds from cadence.
Reconciled attribution

Each sale credited once.

Touchpoints reconciled to a single credited conversion, so the platform that drove the margin is the one that gets the credit.

Method. Deterministic plus survey signal, blended.
Recommendations that close the loop

Ranked actions, then measured.

Every surface produces ranked actions with the margin at stake, and the engine checks whether the number actually moved after you act.

Method. Margin-ranked action queue with outcome verification.
Natural-language insight

A plain-English read, every week.

Investigator agents pull the data, find the drivers, and the model writes up what happened, why, and what to do next.

Method. Multi-investigator synthesis, written by the model.
How our AI thinks

Rigorous where it counts, honest where it must be.

The difference is not a flashier chart. It is models you can trust, fit to your business and tied to your numbers.

Learns your business, not a template

Models are fit per account from your own data. There are no hardcoded rules written for an industry, so the read reflects how you actually trade.

Reconciled to source

Every input a model uses ties back to the order, cost or entry behind it. The intelligence inherits the same lineage as the numbers.

Honest by default

When a feed or a model is not ready, the surface says so. You see an awaiting-data state, never a fabricated number or a false precision.

Closes the loop

Recommendations log the action taken and then measure whether the margin moved. The engine learns from what worked.

The methods behind it

Real techniques, not a black box.

We are happy to show our working. These are the methods doing the modelling, the same ones an analyst would defend in a review.

Forecasting & demand
Holt-Winters smoothingP10 / P50 / P90 bandsDriver-based forecastsPrice elasticity (log-log OLS)
Signals & detection
Pearson lag correlationBollinger bandsZ-score detectionU-shaped extrema models
Customers & value
RFM percentile scoringK-means clusteringCM1 lifetime valueCohort and Gini concentration
Custom Business Intelligence extends this with bespoke models built for your question, on the same engine.
See the intelligence run on your data.A 30-minute walkthrough on your numbers, not a generic demo.