NeuraOrderHow we test NeuraOrder
Every claim on this site comes from replaying real retail sales, day by day, through a simulated stockroom. NeuraOrder only ever sees the sales up to each day it makes a decision. Here is the method, the full results and their limits.
Saved on 1,000 products in 26 weeks
48% lower cost of lost sales and carrying stock than a common reorder rule.
Fewer weeks out of stock
2.9% of weeks instead of 12.6%.
Less fresh-food waste
With shelf life set, on two samples of 400 grocery products.
Honest likely ranges
Real sales landed inside the 80% range 85% of the time.
Fewer stockouts at about half the cost
Two years of real daily sales from a UK online retailer (UCI Online Retail II). We replayed the last 26 weeks for 1,000 products that were never used to tune NeuraOrder. Each product had a 14-day supplier lead time and weekly ordering, and sales lost to empty shelves stayed lost. Both approaches faced identical demand and started from the same stock.
Where the money went
Cost of lost sales plus carrying stock, 1,000 products over 26 weeks
- Sales lost to stockouts
- Cost of carrying stock
NeuraOrder spent more on carrying stock (it held 50 days of supply, the rule 22) and far less on lost sales.
Data: UCI Online Retail II (Chen 2012, CC BY 4.0). The last 26 weeks of real daily sales for 1,000 products, replayed through a simulated stockroom: 14-day supplier lead time, weekly ordering, unmet demand lost, 50% margin and 25% yearly holding cost assumed. The common reorder rule orders up to 1.5× the last 28 days' average sales for the next 21 days.
Fewer stockouts, more sales served
Same 1,000 products, same 26 weeks
- Common reorder rule
- NeuraOrder
Stockout weeks: 0–15% scale. Demand served: 0–100% scale.
Data: UCI Online Retail II (Chen 2012, CC BY 4.0). The same replay as the cost chart: 14-day lead time, weekly ordering, unmet demand lost. The common reorder rule orders up to 1.5× the last 28 days' average sales for the next 21 days.
| Result over 26 weeks | Common reorder rule | NeuraOrder |
|---|---|---|
| Weeks with a stockout | 12.6% | 2.9% |
| Customer demand served | 82.0% | 93.7% |
| Cost of lost sales + carrying stock | $189,200 | $97,700 |
| Days of supply held | 22 | 50 |
The rule orders up to 1.5× the last 28 days' average sales for the next 21 days, a common rule of thumb. NeuraOrder held more stock than the rule where running out was costlier than carrying it: at the assumed 50% margin and 25% yearly holding cost, a lost sale costs far more than a month on the shelf. It sizes orders from each product's own odds, so stores with thinner margins get leaner stock.
Forecast ranges were honest too: actual sales landed inside NeuraOrder's 80% likely range 85% of the time.
Less waste when products expire
Real daily sales of fresh groceries, with each day's stockouts recorded (FreshRetailNet-50K). We replayed 5 weeks with a 2-day lead time, orders every 3 days, and stock that spoils after 6 days. We used two separate samples of 400 products: one to explore, and a second held back to confirm the results.
Fresh food thrown away
Share of units bought that spoiled, two separate samples of 400 products
- Without shelf life
- With shelf life set
About 40% less waste in both samples. Profit rose 7.0% and 4.9%, while 95.9% and 96.8% of demand was still served.
Data: FreshRetailNet-50K (Wang et al. 2025, arXiv:2505.16319, dataset by Dingdong-Inc, CC BY 4.0). Real daily sales replayed for 5 weeks through a simulated stockroom: 2-day lead time, ordering every 3 days, stock spoils after 6 days. Both bars are NeuraOrder, ordering with and without each product's shelf life.
| Result | Without shelf life | With shelf life set |
|---|---|---|
| Units wasted (sample 1) | 11.6% | 7.0% |
| Units wasted (sample 2) | 8.3% | 4.7% |
| Profit change | — | +7.0% / +4.9% |
| Demand served | 97.6% / 98.2% | 95.9% / 96.8% |
Knowing the shelf life means ordering only what's likely to sell before it expires: a little less demand served, much less thrown away, more profit.
Savings estimates that hold up
On real sales, we simulated suppliers offering volume discounts or announcing a 10% price rise. NeuraOrder's planner suggested bulk buys only where they'd still save money if sales ran slow. We then compared what it predicted with what the replay actually saved.
Estimated savings vs real savings
Bulk buys at the default (cautious) setting: 63 and 10 products bought ahead
- Estimated before buying
- Real savings in the replay
Volume-discount savings came in 3% below the estimate; price-rise savings 12% above it. Each panel has its own scale.
Data: UCI Online Retail II (Chen 2012, CC BY 4.0), 167 products not used in the other tests, with simulated supplier prices: 5% off orders of 4+ weeks of demand and 10% off 8+ weeks, or a 10% price rise announced 28 days ahead, replayed through the same simulated stockroom (14-day lead time, weekly ordering).
| Scenario | Realized ÷ predicted savings | Bulk buys that lost money |
|---|---|---|
| Volume discounts | 0.97 | 2% |
| Announced 10% price rise | 1.12 | 0% |
What these results do and don't mean
- Out of sample. On every simulated day, NeuraOrder only saw sales up to that day.
- No peeking. Settings were tuned on one set of products and the results reported on a separate set it had never seen.
- Same conditions for everyone. Every approach faced identical demand, lead times and starting stock.
- The AI. NeuraOrder combines a pretrained AI forecasting model with statistical methods. It never trains on your data.
- Simulated, not your store. The sales are real and public, but the stockroom, costs and supplier prices are simulated. Real lead times vary, and your margins, products and suppliers will change the numbers. These are test results, not a guarantee.
- See it on your own data. Inside the app, the Forecast accuracy page shows how forecasts would have done on your store's past sales.
Data credits
Chen, D. (2012). Online Retail II [Dataset]. UCI Machine Learning Repository. doi.org/10.24432/C5CG6D. Licensed under CC BY 4.0.
Wang, Y. et al. (2025). FreshRetailNet-50K: A Stockout-Annotated Censored Demand Dataset for Latent Demand Recovery and Forecasting in Fresh Retail. arXiv:2505.16319. Dataset by Dingdong-Inc, licensed under CC BY 4.0.
The stockroom simulation, costs and supplier prices are our own, added on top of the real sales.
See what it does with your sales
Every plan starts with a 30-day free trial, and the accuracy page shows how NeuraOrder does on your own history.