Are Forecast Accuracy and Plan Stability the Same Thing? — Three Structural Reasons Why a 10-Point Drop in MAPE Doesn’t Fix Workshop Chaos

A home goods company spent most of 2025 implementing a demand forecasting system, improving the monthly forecast accuracy at the SKU level from 62% to 74%. Both the sales and planning departments were satisfied. However, by the end of the year, the on-time delivery rate only increased from 81% to 83%, and the workshop still had to work overtime, with urgent orders causing disputes as usual. The planning manager's original words were: "The forecast is accurate, but the plan is still chaotic."

This is not an isolated case. Any company that treats "plan chaos" as a "forecast inaccuracy" issue will end up in the same situation: money spent on models and data, symptoms persisting in the workshop and delivery, with no alignment between the two.

Conclusion first: Forecast accuracy and plan stability are two different things, separated by three structural gaps—discrepancy in granularity between forecasting and scheduling, lack of constraints in the plan, and no cost for plan changes. Without addressing these gaps, a 10-point improvement in forecast accuracy will only make an unexecutable plan more precise.

1. Distinguish Between the Two Metrics, Don’t Use MAPE to Manage Plans

Forecast accuracy measures whether the predicted numbers match the actual demand, while plan stability measures whether the commitments are met. The former is input quality, and the latter is output reliability. Whether the execution layer trusts the plan depends on the latter, not the former.

The third row in this table is often the most overlooked. Criterion: If the plan adherence rate is below 70%, the problem lies in the plan's executability, not the forecast. At this point, continuing to optimize the forecast model is like waxing a wheel that won't turn.

Another simpler criterion: Look at the sign of the forecast bias. If the bias is consistently in the same direction for three consecutive months (e.g., overestimating by more than 10% each month), it's not a model issue, it's an incentive issue—sales forecasts are tied to sales targets, making optimistic biases rational. In such cases, changing the algorithm won't help; what needs to change is "who makes the forecast and who bears the bias."

2. Reason One: Granularity Gap—Aggregated Layer is Accurate, Detailed Layer is Chaotic

Demand forecasting is typically done at the "product family × month" level because this level has historical data and statistical significance. However, scheduling requires "SKU × process × day." Moving from the former to the latter involves two amplifications:

• Structural drift within the category: The total is correct, but SKU A sells 30% more, and SKU B sells 30% less, with the total offsetting each other. The total accuracy rate might be 90%, but the SKU-level accuracy could be only 50%.
• Misalignment in time distribution: Monthly demand of 3000 units does not mean 100 units per day. Industries with concentrated shipments at the beginning of the month and end-of-month surges will inevitably face stockouts in the first ten days and excess inventory in the last ten days if the demand is averaged daily.

Passing Line: The SKU-week accuracy rate for Class A products (the 20% that account for 70% of sales) should not be less than 70%, while Class C products can rely on inventory to cover. If the accuracy rate for Class A products is only 40%, it suggests that you are using the aggregated layer's accuracy to self-comfort.

Counterexample: An electronics company evenly distributed the monthly forecast over 22 working days, resulting in a 62% on-time delivery rate for the first 8 days of the month and 91% for the last 8 days. They assumed it was due to insufficient capacity and added overtime, but the problem remained—because the demand was not evenly distributed, and the plan smoothed it out.

The correct approach is not to "make the forecast more detailed," but to layer it: the operating level (S&OP) manages totals and resources by product family and month; the planning level (MPS) manages inventory and capacity balance by SKU and week; the execution level manages process sequence on a daily rolling basis. Layers should be connected by rules, not by a single table that spans all levels. Which categories are worth more detailed forecasting can be directly referenced from the six passing lines in What Categories Are Worth AI Demand Forecasting? (16.3.1)—categories with insufficient data preparation will only produce noise if forced into detailed forecasting.

3. Reason Two: No Constraints in the Plan—That’s Not a Plan, It’s a Wish

The second gap is more hidden: many companies' "plans" are simply the forecast numbers passed down without considering capacity, material availability, or changeover time.

A plan without constraints is inherently unexecutable. An unexecutable plan leads to a chain reaction:

1. The workshop finds that it cannot complete the plan and adjusts the sequence on its own.
2. The adjusted reporting data becomes disconnected from the plan, making actual capacity and cycle times unmeasurable.
3. The next round of planning is based on distorted base data, leading to even larger errors.
4. The planning department loses credibility, and sales start to directly contact the workshop, placing urgent orders via phone.

Once this cycle forms, optimizing the forecast only accelerates it: the more accurate the forecast, the more authoritative the plan appears, the more unexecutable it becomes, and the more hidden the execution layer's adjustments become.

To determine whether the plan includes constraints, look at three factors: whether changeover time is calculated during scheduling, whether key material availability is verified, and whether it is subject to minimum batch size limits. If none of these are considered, the plan is essentially a "demand list."

Passing Line: The material availability verification should cover 100% of Class A materials before scheduling, changeover time should be included in the scheduling model, and key process capacity utilization should be between 85% and 95%. Below 85% is capacity waste, and a plan that consistently exceeds 100% relies on overtime and urgent orders, which is equivalent to not scheduling at all. The consequences of incomplete constraints are detailed in The Real Boundaries of Intelligent Scheduling (16.3.2).

4. Reason Three: No Freeze Period—Changes Are Free, So Plans Lack Credibility

The third gap is a governance issue, not a technical one.

If sales can place urgent orders at any time, the boss can change priorities at any time, and changes do not require anyone to bear the cost, the optimal strategy for the plan is to "re-schedule every day." In response to a daily changing plan, the execution layer's only rational action is self-protection: overproducing semi-finished goods, reserving capacity margins, inflating reported data, and hiding the true progress.

This creates a paradox: The more flexible the plan, the less transparent the execution; the less transparent the execution, the more the plan relies on follow-up.

To break this cycle, changes must be priced. The approach is to divide changes into three categories:

Criterion: The proportion of changes within the freeze zone. If it exceeds 10%, the freeze period is ineffective; a healthy value is below 5%. Another criterion is the distribution of change reasons—if more than 60% of changes are due to "customer last-minute requests," it's a sales commitment management issue; if it's due to "materials not arriving," it's a procurement and planning coordination issue. The remedies for these two causes are entirely different, and mixing them will only lead to mutual blame.

Expected Interval: By only setting the freeze period and change thresholds, without touching the forecast model, the plan adherence rate usually improves by 15～25 percentage points within 3 months, and the weekly change rate drops below 10%. This is the gap with the smallest investment and the fastest results.

5. Diagnosis Sequence: Measure Stability First, Then Fix Granularity, and Finally Discuss the Model

Many companies start by buying forecasting software, which is the wrong order. The reasonable sequence is:

1. First Week: Measure Three Numbers. Plan adherence rate, weekly change rate, and SKU-week accuracy rate. Once these three numbers are laid out, the problem layer becomes immediately clear—low adherence and high accuracy indicate issues with constraints and changes; low accuracy and acceptable adherence indicate issues with granularity.
2. First Month: Set Change Thresholds. Define the freeze period, signatories, and include changes in performance evaluations. This step does not require a system, just a rule table and someone willing to sign off.
3. Second and Third Months: Add Constraints. Incorporate changeover time, material availability verification, and minimum batch size into the scheduling logic. This usually involves modifying configurations in the existing system or adding a scheduling tool.
4. From the Fourth Month On: Discuss the Forecast Model. At this stage, the marginal benefits of forecast accuracy can truly be absorbed by the plan.

The underlying judgment behind this sequence is: The impact of forecast accuracy on delivery is marginally decreasing, while the impact of plan stability on delivery is stepwise. An accuracy improvement from 55% to 65% might only move delivery by 1～2 points; a plan adherence rate improvement from 60% to 85% often jumps by more than 10 points. The reason is that the former improves information, while the latter improves execution.

However, forecast errors cannot be eliminated, only absorbed. Absorption relies on two things—safety stock and capacity buffer. Therefore, what truly needs to be calculated is not "can the accuracy rate be improved by another 3 points," but "at the current error level, where is the cheapest place to set the buffer?"

The delivery delays caused by plan chaos mostly occur at departmental handoffs, not in the workshop, as detailed in Seven Breakpoints from Order to Delivery (16.2.1). The root cause of why three scheduling systems are still reliant on Excel lies in the data premises of the plan, as discussed in Three Scheduling Systems Later, Why Still Rely on Excel? (16.2.2).

In a Nutshell

Forecast accuracy and plan stability are separated by three structural gaps: granularity (aggregated layer is accurate, detailed layer is chaotic), constraints (plan lacks capacity and material considerations), and change cost (urgent orders are free). The sequence for fixing these gaps is to measure adherence first, set the freeze period next, add constraints, and finally adjust the model—reversing this sequence will ensure that money is spent where it has no visible effect.

This article belongs to the "Digitalization and AI Zone · Operational Digitalization" direction, targeting operational, planning, and digitalization leaders, and does not involve quality management system clauses.

Knowledge code: 16.2.2 Version: v20231010 Author: QTank