We released an AI-powered forecasting engine in early March. It was a big improvement. Then we kept going.
The original model learned from your historical data and cut over-prediction dramatically — from the 20–34% range down to under 7%. But the approach had room to grow: it was still treating all products similarly, it didn’t fully capture how sales patterns shift across a multi-day event, and it had no mechanism for handling unusual outlier days without letting them distort future predictions.
The new engine — Enhanced Forecast v3 — is a complete rebuild that addresses all of this.
How it works
Rather than grouping sales into historical buckets by warehouse rank and day type, v3 trains a machine learning model on your specific sales history. It learns the actual patterns in your data: which warehouses consistently outperform their rank, how your best-selling products behave differently from your slow movers, which weeks of the year tend to be strong or soft, and how sales typically trend across the arc of a long event.
The model uses an ensemble approach — it combines a gradient-boosted learner with a linear model. The practical effect is that it captures both the complex, non-obvious patterns and the stable long-run trends. And when an exceptional sales day happens — a product inexplicably doubles its usual volume — the model handles it as an outlier rather than updating its expectations dramatically based on a single data point.
The accuracy numbers
We tested both the old and new engines against the same vendor’s complete sales history: over 15,000 actual day-location-product data points from a full year of events. Each engine generated forecasts without seeing the actual outcomes, and then we compared.
| Metric | Classic Forecast | Enhanced v3 | Improvement |
|---|---|---|---|
| Avg. units off per day/location | 1.99 | 1.25 | 37% more accurate |
| Average % error | 79.6% | 51.6% | 35% better |
| Sensitivity to big misses | 3.47 | 2.09 | 40% better |
| Full-year volume: forecast vs. actual | 57,420 vs. 47,628 | 46,454 vs. 47,635 | Less than 3% off |
The last row is worth dwelling on. Over an entire year of events, the old engine called for 9,792 more units than actually sold. The new engine missed total volume by 1,181 units — across tens of thousands of individual forecasts. That kind of accuracy at scale means your inventory plans, staffing goals, and replenishment timing are all working from numbers that reflect reality.
What changes for you
Staffing goals become achievable. When your daily targets are set based on what’s actually likely to happen — not a historic average that skews high — your team’s performance can be measured against something fair.
Inventory planning improves. If the forecast knows a product tends to move fast on day one and slow down by day three, you can time replenishment accordingly instead of staging everything at the start.
It keeps learning. The model retrains regularly on your latest data. Every event you run adds to what it knows about your specific operation.
How to switch
Go to Forecast Configure in your ZenShows dashboard. In the Algorithm dropdown next to your retailer, select Enhanced Forecast v3. It saves automatically. Hit Rebuild All and your next forecast run uses the updated model.
Your locked event goals aren’t touched. Only unlocked events get new predictions. And if you want to see how the two engines compare for your specific account, the Algorithm Performance panel at the bottom of the Forecast Configure page shows the side-by-side accuracy data for your own history.