What Is a Good Forecast Accuracy by Industry?

A good forecast accuracy depends on what you sell and how lumpy demand is, not on how smart your planners are. For a high-volume food manufacturer, 80-90% at the item level is realistic and healthy. For a $40,000 capital-equipment SKU, hitting 50% is a strong result, and demanding 90% is a fantasy that wastes everyone's time. The single biggest mistake I saw running demand planning at a $250M industrial manufacturer was leadership demanding one number across the board without knowing whether it was generous or insulting for a given product family.

So let me give you the numbers that actually matter. And the trap built into the question itself.

Forecast accuracy is meaningless without three pieces of context

Before you judge any accuracy figure, you need three things. Without them, the percentage is noise.

• Aggregation level. Accuracy at the total-company level is always higher than at the SKU-location level, because errors cancel out. A planner reporting "92% accurate" at the national monthly level might be running 60% at the DC-week level where replenishment actually happens.
• Time bucket. Weekly is harder than monthly. Monthly is harder than quarterly. Same forecast, different scoreboard.
• Demand variability. This is the one most teams skip, and it's the most important.

If someone quotes you an accuracy number and can't name the aggregation level, time bucket, and variability band, treat it as marketing.

How forecastable is the product in the first place?

You can't out-plan physics. The forecastability of a product is largely fixed by its demand pattern, and the standard way to measure that comes from the Syntetos-Boylan classification (2005), which sorts demand using two numbers: average inter-demand interval (ADI) and the squared coefficient of variation of order sizes (CV²).

The scheme splits products into four buckets using cutoffs of 1.32 for ADI and 0.49 for CV²:

Smooth demand is regular in both timing and quantity, so it's easy to forecast. Lumpy demand swings in both, and as one method-selection review in Sādhanā (2020) puts it, it's extremely hard to produce a reliable forecast for. Know which bucket a SKU lives in before you set a target for it.

Forecast accuracy benchmarks by industry

Here's a realistic range for SKU-level, monthly forecast accuracy, measured as 100 − WMAPE and weighted by volume. These are operator numbers, not vendor-brochure numbers.

Notice the spread. A 60% in industrial B2B can be genuinely strong. The same 60% in food means the planner is asleep. If you want the manufacturing-specific cut of these numbers, I broke them down further in our forecast accuracy benchmarks for manufacturers piece.

One more warning on the table. Don't read a row as a hard target for every SKU inside that industry. A food manufacturer's seasonal limited-edition runs at apparel-grade unpredictability, and a heavy-machinery maker's consumable parts can forecast as cleanly as CPG. The industry sets the center of gravity. The demand pattern of the individual item sets where that item actually lands.

Why MAPE lies to you

Most teams report MAPE (Mean Absolute Percentage Error) and call 100 − MAPE their "accuracy." It's the default in every ERP. It also breaks in the exact situations that hurt you most.

• MAPE blows up on low-volume SKUs. Forecast 2 units, sell 1, and your error is 100%. That one item can wreck a category average even though the dollar impact is trivial.
• MAPE is undefined when actuals are zero. As Hyndman and Koehler (2006) document, the metric is degenerate whenever an actual hits zero, so intermittent items get dropped or fudged.
• MAPE rewards over-forecasting. The downside error is capped at the actual; the upside is uncapped. Bias creeps in unnoticed.

What to measure instead

Hold your team to metrics that survive real data:

• WMAPE (weighted MAPE) — weight by volume or revenue so the metric reflects the P&L. It's also defined when individual actuals are zero, which MAPE is not. We compare the two directly in MAPE vs WMAPE.
• Bias (mean error) — track it separately from accuracy. You can be accurate and biased, or unbiased and inaccurate. Persistent positive bias is how stranded inventory gets built.
• MASE for intermittent items — the scaled error from Hyndman and Koehler's intermittent-demand paper (2008) compares your forecast against a naive baseline, which is the only fair test on spare parts.

What "good" means for a CFO, not a planner

Here's the reframe I wish I'd had earlier. Forecast accuracy is an input, not an outcome. The CFO doesn't care about MAPE. They care about three things downstream of it.

1. Stranded inventory — cash tied up in stuff that won't move at full margin.
2. Service level and fill rate — revenue you didn't lose to stockouts.
3. Expediting and obsolescence cost — the tax you pay for being wrong.

The cash at stake is real. Inventory carrying cost runs 20-30% of inventory value per year, so a few points of accuracy on a smooth A-item can free seven figures in working capital. The same few points on a long-tail C-item is rounding error. Chase accuracy where it converts to cash.

Here's the math that makes this concrete. Suppose better accuracy lets you trim $4M of excess stock across your A-items. At a 25% carrying rate, that's $1M of recurring annual cost you stop paying, plus a one-time $4M cash release back into working capital. Now run the same improvement on a tail of slow-moving C-items holding $200K total. The annual saving is $50K, and the planner hours to get there cost more than that. Accuracy is not free to chase, so spend the effort where the dollars are.

A practical target-setting framework

Don't set one number. Segment, then set a floor per segment.

• A-items, smooth demand: target 85%+ and treat anything under 80% as a problem.
• A-items, variable demand: target 70-80%, and invest in causal drivers (promo, price, weather).
• B/C-items, intermittent or lumpy: stop chasing point accuracy. Manage with safety stock and service-level targets instead.

The cleanest way to slice this is an ABC-XYZ inventory analysis, which crosses dollar importance (ABC) with variability (XYZ) so your floors line up with both value and forecastability.

Then measure whether your planners actually help

Track forecast value added (FVA): does a planner's manual override actually beat a naive statistical baseline? Most teams have never run this test, and the results are sobering. A study of supply chain companies published in the International Journal of Forecasting (2024) found that 52% of their forecasts were worse than a simple random walk. Over half the time, all that planning effort made the forecast worse than doing nothing.

That's the most expensive thing in the room, and almost nobody measures it. I walk through the mechanics in our forecast value added how-to.

Where AI moves the number, and where it doesn't

AI gets sold as a cure for every accuracy gap. It isn't. But on the right demand patterns, the lift is real and well-documented.

McKinsey (2023) reports that AI-driven forecasting can cut errors by 20-50% and reduce inventory by 20-30%. Gartner (2025) predicts 70% of large organizations will adopt AI-based supply chain forecasting by 2030. The catch: those gains land mostly on smooth and erratic demand with rich causal signals.

On genuinely lumpy, deal-driven capital equipment, a transformer model won't out-forecast a human who knows the sales pipeline. Point your AI investment where the demand pattern is forecastable and the dollar volume is high. Spend the rest of your energy on safety stock and service-level design.

The bottom line

A good forecast accuracy isn't a single magic number. It's a segmented set of floors, measured with the right metric, weighted by dollars, and judged by whether it frees working capital and protects service.

If your team reports one company-wide accuracy figure and calls it a day, you don't have a forecasting problem yet. You have a measurement problem that's hiding one. Fix the measurement first, then the forecast, then the cash.

Frequently asked questions

Is 80% forecast accuracy good?

It depends entirely on the product. For high-volume food, CPG, or pharma at the SKU-month level, 80% is solid and expected. For industrial B2B, spare parts, or capital equipment, 80% would be exceptional or even implausible, because the underlying demand is lumpy and far less forecastable.

What is the best metric for measuring forecast accuracy?

For most product portfolios, use WMAPE (weighted MAPE), since it weights error by volume or revenue and stays defined when individual actuals are zero. For intermittent items like spare parts, use MASE, which scales your error against a naive baseline. Always track bias separately, because a forecast can be accurate on average yet consistently over- or under-shoot.

Why is MAPE a poor accuracy metric?

MAPE is undefined when actuals are zero, explodes on low-volume SKUs where a one-unit miss reads as 100% error, and quietly rewards over-forecasting because its downside error is capped while the upside is not. Hyndman and Koehler showed it's degenerate in common situations and recommended scaled metrics like MASE instead.

How does demand variability affect a realistic accuracy target?

Variability sets the ceiling on how forecastable a product is, regardless of planner skill. The Syntetos-Boylan classification uses inter-demand interval and the coefficient of variation to sort items into smooth, erratic, intermittent, and lumpy buckets. Smooth items support high point-forecast targets; lumpy items are better managed with safety stock than with a tighter forecast.

Does AI improve forecast accuracy enough to be worth it?

For high-volume products with smooth or erratic demand and strong causal signals, yes. McKinsey reports AI can cut forecast errors by 20-50% and inventory by 20-30% in those conditions. For lumpy, low-volume, deal-driven items, the lift is marginal, and human pipeline knowledge usually beats the model.