How Auxetic forecasts
Three forecasts. One consensus you can defend.
Auxetic runs three diverse forecasting approaches against your data, has them critique each other, and gives you the consensus — and the disagreements — so you know which forecast to trust, when, and why.
The panel
Three diverse forecasts. Not three runs of the same model.
Diversity is the point. Each source has different blind spots, so the disagreements between them are themselves a signal.
Classical
AutoARIMA / AutoETS / Theta
Strong for stable trend and seasonality. Decades of evidence behind them. Honest about what they can't see.
Machine learning
Lag-feature gradient boosting
Strong for regime changes, recent shifts, and exogenous signals. Captures interactions the classical models can't.
Foundation forecaster
Pretrained time-series model
Strong for short histories and zero-shot transfer. Brings priors from millions of other series.
Adversarial step
Each one challenges the others.
Three independent analyst personas read the data, every forecast, and the backtests — then critique what the others got wrong. The consensus is the part that survived disagreement.
Statistician
Stationarity, seasonality, regime breaks, interval calibration. Will call out an ARIMA fit that smuggled a unit root.
ML practitioner
Lag structure, recent residuals, leakage, drift. Will catch the gradient-boosted forecast that's hugging the last data point.
Operator
Business context, holidays, events, structural shifts. Will flag a forecast that ignores the fact you launched a new SKU last month.
Synthesis
One forecast. With the disagreement preserved.
The synthesizer takes all three forecasts and all three critiques and produces a single consensus — without throwing away the parts you should weigh.
Weighted-ensemble point forecast
Sources weighted by their rolling-backtest accuracy and adjusted by the panel's critique.
Intervals that reflect model disagreement
Not just statistical noise — the variance across the panel widens the interval honestly when models diverge.
Agreement / disagreement summary
“All three agreed within X%” or “Statistician dissented because Y” — in plain English.
Dissent preserved
The minority forecast is kept as a caveat the user can weigh, not silently discarded.
Transparent confidence
Forecasts come with how much to trust them.
Every Auxetic forecast ships with explicit confidence signals — not just a single interval and a vibe.
Panel-agreement score
How tightly the three forecasts converge. A wide spread is a real, surfaced signal that you should trust the forecast less.
Backtest accuracy per source
Rolling-window MAPE / sMAPE for every source, plus the baseline (seasonal naive). You see exactly which model earned its weight.
Interval coverage
Does the 80% interval actually cover 80% of past actuals? We measure and show it — not just assume it.
Honest caveats
Short history, irregular sampling, regime shifts, missing exogenous data — surfaced as plain-English warnings before you act.
Outliers, surfaced
The data points that don't fit, called out.
Outliers can be the most valuable signal in the data — or the biggest source of bad forecasts. Auxetic finds them, names them, and lets you decide what to do.
Data-quality outliers
Anomalous rows flagged at profile time — statistical (z-score, IQR, Hampel) and rule-based (suspicious zeros, negative revenue, future dates).
Forecast-residual outliers
When the backtest got it wrong, exactly where — surfaced as 'moments your model didn't see coming.' Often the most valuable signal in the data.
Recommended action
For each outlier: ignore (one-off), label (recurring), or treat as a regime change. You decide; the forecast adapts.
Status
What ships today. What's next.
Today
- Auto-discovered KPIs and a data-readiness score on every upload.
- Backtested forecast with prediction intervals and a model leaderboard.
- Champion-vs-baseline margin rule and model caveats in plain English.
- Numeric-grounding guardrail on the analyst's answers.
Next — in active development
- Three-source panel (classical + ML + foundation forecaster).
- Persona critique step (statistician / ML practitioner / operator).
- Consensus synthesizer with disagreement-aware intervals.
- Panel-agreement score and dissent preserved as a caveat.
- Data-quality and forecast-residual outliers with recommended actions.
We don't put vapor on the marketing site. This page describes how Auxetic forecasts — today and on the near roadmap. If you want to design-partner the next step, talk to us.
Ready to stop guessing?
Join early access and see how Auxetic turns your sales, demand, cash-flow, and operational data into predictive clarity.
We'll invite selected businesses for a Forecast Readiness Audit.