Demand Forecasting Compared: Moving Average, Exponential Smoothing and Machine Learning

There is no "best algorithm", only the best fit

Forecast quality drives everything downstream — safety stock, reorder points, order quantities. The most common mistake in practice is chasing a single advanced algorithm across the whole catalog. With thousands of SKUs and uneven data quality, tiered method selection by SKU characteristics beats one complex global model every time.

Four families of methods

• Simple / weighted moving average — best for stable demand without trend or seasonality. Transparent, explainable, nearly zero maintenance, but slow to react to trend shifts. A multi-window weighted average (e.g. last 7 days, last 30 days, yesterday, 3-month daily average with tuned weights) is the most practical baseline for replenishment because buyers can see exactly where every number comes from.
• Exponential smoothing — SES for flat series; Holt adds a trend term; Holt-Winters adds seasonality for clearly cyclical goods. Parameter-light and automatable. The trap: e-commerce series are full of promotion pulses, and fitting raw history extrapolates a flash sale as "normal demand". Promotion days must be excluded or modeled separately.
• Croston and variants — built for intermittent demand ("0, 0, 3, 0, 0, 1, 0 ..."). It splits the problem into demand interval and demand size, smoothing each before recombining. For long-tail office supplies or spare parts this is far more reliable than forcing a regular method onto a sparse series.
• Machine learning — gradient-boosted trees and similar models work at scale, but only if you have long clean history, feature data (price, promotions, traffic), and a team to maintain feature pipelines and retraining. Missing two of those three, ML delivers a black box and an ops burden instead of accuracy.

Tiered selection guidance

1. A-class (top movers): start with Holt-Winters; evaluate ML only when features and team are in place.
2. B-class (mid-tail): multi-window weighted moving average plus manual coefficient tuning — best value.
3. C-class (long tail, intermittent): Croston, or skip forecasting entirely and replenish by turnover-day caps and minimum order quantities.
4. New / promotional items: keep them out of statistical models; use analogous-product curves and manual plans.

Feed the error back

Forecast error (MAPE, bias direction) should be persisted and reviewed alongside the forecast itself. A consistently over-forecasting model is more dangerous than a crude but unbiased one — it systematically manufactures overstock. Every suggestion should stay traceable back to its data and formula, so error can be located and corrected.