Nori V1 — Replaces XGBoost

Free the Working Capital Your Forecasts Are Tying Up

Synthefy is a foundation model for supply chain demand forecasting, built on a time-series platform deployed at Samsung, NetApp, Deutsche Telekom, and the US Army. See more at synthefy.com.

Download PDFPDF available for download · view online or offline

01 The Problem

Three failures show up in every demand forecast we audit:

  1. The model can’t see what moves your demand. Blind to tariffs, commodity prices, and demand shocks, it stays wrong on the 30 to 50% of your catalog that’s lumpy or seasonal.
  2. No insight into why the forecast moved. Planners are told the number changed, not the reason they need to make a better call and defend it.
  3. Overrides change the number, never the reasoning. Signal and bias get patched in by hand and lost, so the same miss compounds.

The result is about 5% of your lumpy-SKU inventory sitting idle as excess safety stock, which for a $300M company is roughly $1M of working capital tied up and $250K a year in carrying cost.

02 Why Synthefy

Three things Synthefy does that most models can’t.

Accuracy on the SKUs your model gets wrong.

SKU-specific external signals lift accuracy on the half of your catalog that was stuck.

Forecast accuracy, lumpy SKUs12 weeks
40%60%
Incumbent ~42%Synthefy ~60%
Value~5% of lumpy-SKU inventory released, about $1M one-time for a $300M company.

Forecast trust your team can defend.

Every forecast ships with a signal-by-signal breakdown plus a confidence score.

Forecast attribution · SKU-482712.3K units
History54%
Copper20%
Tariffs14%
News8%
Confidence0.84
ValueThe carrying cost you stop paying on the inventory you free, about $250K a year for a $300M company.

A model that learns from your whole team.

Each override teaches the next forecast. The loop compounds.

Override → model feedbackPer planner
ForecastPlanner editModel updateBetter forecast
ValueThe model gets smarter each cycle: accuracy climbs, bias drops, and manual adjustments fall.

Methodology Illustrative for a mid-market company at $300M revenue. Scales with your inventory load and team size. McKinsey, APQC, ISM, BLS.

03 How Synthefy Compares

Synthefy works on top of the tools you already run. Here is what it adds.

Synthefy augments your current stack with a sharper forecasting layer, your existing tools stay in place.

Your dataERPSynthefyforecastingYour existing solutionBlue Yonder, Kinaxis, StockIQDownstreamlogistics, warehouse
CapabilitySynthefyBlue YonderKinaxisStockIQExcel
Time to first forecastHours, API-native12 to 24 monthsMonthsWeeks to monthsDIY
Finds which connected signals causally drive each SKU
Finds cross-product correlations across the whole catalog automatically
Per-forecast numerical signal attribution with a confidence score
Turns each planner override into a training signal
Cost$$$$$$$$$$$$$$$
Built-In Limited Not available

Synthefy works from the signals you connect, and only where external forces actually move demand.

We don’t promise perfect accuracy. We deliver a defensible forecast that significantly outperforms your incumbent stack, with confidence scores planners can act on.

04 Customer Case Study

“Huge under-forecasting on lumpy items. The factory is always lagging behind.”

Head of Supply Chain · Global industrial OEM
The customer
A multi-billion dollar industrial OEM with a global supply chain across six product lines and 160 countries.
The challenge
200,000+ active SKUs and over 1M product-country combinations. Half the catalog lumpy and intermittent, stuck at 40 to 45% accuracy on the incumbent stack.
The approach
Phase zero deployment. Added external signals (copper, commodity ETFs, country macros, news, currency, cross-product), zero-shot.
The outcomes
MetricBaselineWith Synthefy
Accuracy, lumpy SKUs40 to 45%~60%
Time to liveMonths of training and tuning3 days, zero-shot
Working capital releaseBaseline~$145M one-time
Three-year FCF impactBaseline~$255M cumulative

Methodology McKinsey 5% accuracy-to-inventory bridge applied to roughly half of customer’s $5.8B year-end inventory (lumpy SKU portion). APQC and ISM carrying cost benchmarks.

Our model uniquely found some of these correlations

A copper-price signal cut overprediction nearly 6×.

The model found a strong causal link between a key SKU and copper prices — an external input the incumbent never used. Feeding in that single signal collapsed overprediction from +11% to +1.8%.

Cross-product learning caught Turkey’s October spike.

By borrowing demand patterns from similar products, the model detected geo-specific October seasonality the incumbent missed. It predicted 3.6K against 13.8K actual; Synthefy called 12.3K.

What’s next

Bring us your hardest forecast.