T. Brian Jones is co-founder and CTO of Inventory Hero. He leads the engineering behind its Amazon data pipeline, demand forecasting, and the AI platform that lets sellers talk to their live inventory, sales, and supplier data in plain language.
AI demand forecasting uses machine-learning models to predict future sales per SKU, typically blending in seasonality and, where available, signals like price, promotions, and trend. It differs from simple methods (like a moving average) by learning patterns across many products and variables at once. Modern tools often use gradient-boosted models such as LightGBM, which have won large forecasting competitions on rich retail data. The goal is a more accurate demand projection to drive reorder timing and quantity.
Is AI demand forecasting more accurate than traditional methods?
It depends on your data, and the honest answer is nuanced. In the M4 competition, pure machine learning did worse than classical statistical methods, and a hybrid won. In the M5 competition, using Walmart's daily retail data, machine learning swept and every top team used LightGBM, but simple methods stayed competitive at the product-store level and the best entries were ensembles. The pattern: AI wins with many correlated series and rich signals; for a small catalog of lumpy demand, a solid statistical baseline plus buffers is hard to beat.
Does AI demand forecasting work for intermittent or seasonal FBA demand?
Intermittent, lumpy demand (many zero-sales days punctuated by spikes) is genuinely hard for any method, AI included. The standard statistical approach (Croston's method) is biased and ignores trend and seasonality, and no technique reliably solves the problem. AI can help when you have enough history and real signals, but for a typical FBA SKU with a few years of erratic sales, the practical answer is a sensible baseline forecast plus safety stock and lead-time buffers sized for the uncertainty, rather than trusting any single 'accurate' number.
Should I trust an AI forecast to set my reorders?
Use it as an input, not an oracle. A forecast, however it is produced, feeds a decision that also has to respect your lead time, minimum order quantity, available cash, and Amazon capacity and storage constraints. Treat the AI forecast as the demand estimate, then let your reorder logic and your judgment turn it into an order. Be especially wary of vendor claims like blanket 95 percent accuracy; aggregate accuracy numbers hide poor per-SKU performance exactly where your reorder decisions are made.
Read article
AI demand forecasting uses machine learning to predict future sales, but the size of the gain depends heavily on your data. The short version: AI wins most when you have many SKUs and rich signals like price and promotions, for sparse and lumpy demand a solid statistical baseline plus safety stock still captures most of the achievable accuracy, and blanket accuracy claims should be treated with suspicion because aggregate numbers hide the per-SKU reality. Below is what the research actually shows and how to use AI forecasting well.
AI demand forecasting uses machine-learning models to project future sales per SKU, typically blending seasonality and, where available, external signals like price, promotions, and broader trend. The difference from a simple method like a moving average is that a model can learn patterns across many products and variables at once, rather than extrapolating one series in isolation.
Modern tools often use gradient-boosted models such as LightGBM, which have a strong track record on rich retail data. Inventory Hero, for example, uses a LightGBM-based approach with seasonality blending. But a model is only as good as the data and the demand pattern it is fed, which is where the honest story begins.
The best evidence comes from the large, public M-competitions, and it is more nuanced than vendor marketing suggests.
In the M4 competition, pure machine learning methods actually did worse than classical statistical methods, and the winner was a hybrid that combined the two. That was a genuine surprise to the "AI beats everything" narrative.
In the M5 competition, which used Walmart's daily retail sales, machine learning swept for the first time: every one of the top teams used LightGBM, and the ML methods beat the statistical benchmarks. But even there, simple exponential smoothing stayed competitive at the individual product-store level, and the winning entries were ensembles, not a single model.
The pattern across both: machine learning wins when you have many correlated series and rich exogenous signals; classical methods remain strong, cheap baselines, especially for a single erratic series.
Here is the reality for many FBA catalogs. A lot of SKUs have intermittent, lumpy demand: long stretches of low or zero sales punctuated by spikes. This is genuinely hard for every method.
The standard statistical tool for it (Croston's method) is known to be biased and ignores trend and seasonality, and honestly, no technique reliably solves intermittent demand. AI does not have a magic answer here either. For a typical SKU with a few years of erratic history, chasing a more "accurate" point forecast has diminishing returns; sizing safety stock and lead-time buffers for the uncertainty matters more than the forecast's decimal places.
None of this means AI forecasting is hype. It adds real value in specific conditions:
Large catalogs. More SKUs and correlated series are exactly where ML models shine.
Rich signals. If you have price, promotion, seasonality, and demand drivers to feed it, a model can use them; a moving average cannot.
Seasonality. Blending seasonal patterns into the forecast is something models do well and manual methods do crudely.
Scale of attention. A model can forecast thousands of SKUs consistently, where a human would cut corners.
If you have a big, signal-rich catalog, AI forecasting is worth it. If you have a small, steady one, a disciplined baseline may capture nearly the same accuracy for far less complexity.
Seasonality is where the gap shows up in dollars. Say a holiday SKU sells about 60 units a day in the off-season. A flat 30-day moving average going into December still projects roughly 60 a day. A seasonal model trained on the last two Decembers projects 110 a day. With a 30-day lead time, the reorder point moves from about 1,800 units (60 times 30) to about 3,300 units (110 times 30). Trust the flat average and you stock out in your best month; that 1,500-unit gap is the difference a seasonal forecast catches. The forecast is only an estimate, but on a seasonal SKU the estimate you choose visibly changes the order.
Be skeptical of the numbers. Blanket claims like "30 percent improvement" or "95 percent accuracy" are marketing; even reputable ranges (like consultancy figures of 20 to 50 percent error reduction) are ranges across many cases, not guarantees for yours. The deeper trap is that aggregate accuracy hides per-SKU performance: a model can look great on average while being unreliable on the specific products where you actually place orders. Ask how a forecast performs on your important SKUs, not the catalog average.
Use AI forecasting as the demand estimate, especially across a larger catalog with real signals.
Do not trust a single number. Size safety stock and lead-time buffers for the uncertainty around it.
Feed it into a real reorder decision with your reorder point logic, MOQ, cash, and Amazon capacity limits.
Match the method to your catalog. As a rough rule, a catalog of 200-plus active SKUs with a year or more of clean history is where ML earns its keep; under 50 lumpy SKUs, a safety-stock buffer covers most of the gap a fancier forecast would close.
Keep the human on the commitment. The forecast informs; you decide.