Reducing forecast error is less about swapping models and more about fixing what feeds them. Most programs start with the model, the fourth fix, not the first. In the M5 forecasting competition, the machine learning models that won outperformed the strongest statistical benchmarks by more than 20 percent, yet a stronger model still cannot repair demand history that was already wrong before training began.
Reducing forecast error comes down to four fixes, applied in order: repair the demand history a model learns from, separate items that can be forecast from items that structurally cannot, forecast and measure at the level a decision actually needs, and govern human overrides so only the adjustments that beat the model survive. This guide explains each source and the order to fix it in.
How to Reduce Forecast Error: Where It Actually Comes From
Forecast error rarely comes from one place. It comes from four, and they compound.
History that misrepresents demand. Item transitions and stock-outs teach the model a version of demand that never happened.
Structurally unforecastable demand. Some items are lumpy by nature. No model closes that gap.
The wrong level. Forecasting and measuring at the wrong level of granularity can make error disappear on paper while every SKU-location decision stays wrong.
The human overlay. Planner adjustments help on some items and hurt on others, and most teams never separate the two.
Model choice sits inside the second source and is rarely the binding constraint. The rest of this guide works through each source in the order it should be fixed.
Source 1. Repair the Demand History Before Touching the Model
A model is only as reliable as the demand series it learns from. Four issues do most of the damage.
Phase-in and phase-out linkage. When one item replaces another, a system that reads them as two unrelated series sees a collapse followed by a cold start, instead of one continuous demand stream. Both forecasts come out wrong. The fix is an explicit successor link so history transfers from the outgoing item to the incoming one.
Supply-constrained periods. Where an item was unavailable, recorded sales understate demand. Training on that history teaches the model that demand was small. You did not sell less because buyers wanted less. You sold less because there was nothing to sell, and the model cannot tell the difference unless the data does. Supply disruptions of this kind are not rare: the J.P.Morgan Global Manufacturing PMI, compiled by S&P Global, has repeatedly flagged extended supplier delivery times over the past two years, the exact condition that produces censored demand in a raw sales series. The fix is to identify constrained periods, restore the demand that would have materialised, and forecast on the restored series, not the raw one.
Outliers and promotions. Separate one-off spikes from repeatable promotions uplift, and tag promotions rather than smoothing them away. A model that cannot distinguish a promotion from a trend will misread both the next time either occurs.
Returns, transfers, and internal orders. Strip out anything that is not genuine customer demand before it reaches the forecasting engine.
Item coding is what makes the first two fixes possible, and it is worth getting right before any demand planning software is asked to do more with the same messy inputs.
Source 2. Separate Reducible Error From Structural Error
Forecastability is a property of the demand pattern itself, and ignoring it is why so many programs chase accuracy on items that will never deliver it. Not every item deserves the same accuracy target.
Syntetos, Boylan and Croston’s 2005 categorization scheme splits demand into four patterns using the average inter-demand interval (ADI) and the squared coefficient of variation of demand sizes (CV²), with cutoffs of 1.32 for ADI and 0.49 for CV². Kostenko and Hyndman’s 2006 refinement arrived at close to the same split (ADI at 4/3, CV² at 0.5) through a simpler derivation, which is part of why the scheme has held up as a research standard. Petropoulos and Kourentzes’ 2015 extension built forecast combination methods on top of the same categories for regular demand.
| Pattern | What It Looks Like | Realistic Accuracy Expectation | What To Do Instead of Chasing Accuracy |
| Smooth | ADI < 1.32, CV² < 0.49; frequent, steady demand | High; a tighter forecast is genuinely achievable | Keep refining the model here, the effort pays off |
| Erratic | ADI < 1.32, CV² ≥ 0.49; frequent but highly variable order sizes | Moderate; size variability caps how tight the forecast can get | Size safety stock to the variability, not a sharper point forecast |
| Intermittent | ADI ≥ 1.32, CV² < 0.49; infrequent, steady sizes | Moderate; timing is the harder problem, not volume | Track the probability of occurrence, not the quantity alone |
| Lumpy | ADI ≥ 1.32, CV² ≥ 0.49; infrequent, highly variable | Low, by nature, not by modeling effort | Set a buffer and review policy instead of a forecast target |
Forecast accuracy methods that skip this split end up applying one target across a portfolio that behaves nothing like a single pattern. A target of 80 percent accuracy applied evenly across a full portfolio is a target to miss on purpose. A more recent review of machine learning algorithms for intermittent demand, published in the International Journal of Production Research, confirms this classification is still the reference point researchers build from, and finds that machine learning methods add the most value on erratic and smooth series, not on the lumpy end of the spectrum.
Source 3. Forecast and Measure at the Level the Decision Needs
Accuracy improves with aggregation almost by definition. A program that reports error at category or network level can show steady improvement while every SKU-location decision underneath stays just as wrong as before.
Forecast granularity matters here. A forecast built and measured at SKU-location-week resolution answers a different question than one built at category-month resolution, and only one of them matches what a planner is actually deciding. Temporal aggregation is a legitimate technique for sparse items, where forecasting at a coarser time bucket and then disaggregating can outperform forecasting the raw series directly. The rule that matters more than any technique: the level at which you measure error should match the level at which the decision gets made. Model selection and the mechanics of forecasting methods matter less than getting this level right first.
Source 4. Govern the Overlay, Do Not Ban It
Judgmental adjustment is not inherently good or bad. It helps on some items and hurts on others, and most organisations never find out which is which.
A study of more than 60,000 forecasts across four supply-chain companies, published in the International Journal of Forecasting, found that adjustments improved accuracy on average in three of the four companies studied, but positive adjustments (raising the number) were far less likely to help than negative ones, and were made in the wrong direction more often, a pattern the researchers linked to a general bias toward optimism.
Forecast Value Added (FVA) is the discipline that separates useful adjustments from harmful ones. It compares each step in the process- statistical baseline, system output, human override- against a naive forecast as the benchmark, and requires each step to beat the one before it. An adjustment that cannot beat the naive forecast it replaced is not adding value, regardless of the planner’s confidence in it. Governing overrides matters most where S&OP cycles depend on the adjusted number reaching the room, and override governance is covered in more depth in the demand planning software guide.
What AI Actually Does Well, and What It Does Not
The honest answer sits between the two extremes usually offered on this topic. Machine learning models are strong at pattern detection across many demand drivers at once, and at short-horizon demand sensing, both of which fall inside Source 2 and part of Source 3. The M5 competition confirmed that gap: machine learning entries beat the strongest statistical benchmarks by more than 20 percent on the aggregation levels tested.
What AI does not do is repair history you have not tagged, and it does not make lumpy demand forecastable. A model trained on unrepaired, phase-in-distorted, supply-constrained history will simply learn the distortion faster and with more confidence. This is the honest bridge to a planning layer built to automate the repair and classification steps described above, rather than a claim that a new algorithm alone fixes error.
How to Reduce Forecast Error in the Next 90 Days
- Audit the demand history for phase-in, phase-out, and supply-constrained periods.
- Tag successor links and restore demand for constrained periods before any model runs on the data.
- Classify the portfolio into smooth, erratic, intermittent, and lumpy.
- Set differentiated accuracy targets by class instead of one target for the whole portfolio.
- Fix the level at which forecasts are measured so it matches the level the decision is made at.
- Start Forecast Value Added reporting so every adjustment earns its place against a naive baseline.
Metrics That Tell You Which Source Is Failing
| Metric | What It Detects | When It Misleads |
| Bias | Whether the forecast runs systematically high or low over time | Rarely misleads; check this first, since bias points to a process problem, not a model problem |
| MAPE | Percentage error relative to actual demand | Distorts on low-volume or intermittent items, where small absolute errors turn into large percentages |
| WAPE | Percentage error weighted by volume | Can still mask poor performance on high-value, low-volume items that matter operationally |
| Tracking Signal | Cumulative bias relative to absolute error, flags drift | Needs a long enough window; too short a window produces false alarms |
APQC’s Open Standards Benchmarking research on supply chain planning tracks forecast accuracy at the shipping-location level using MAPE as a standard measure, and consistently shows a wide gap between top-performing and median organisations on this metric. If Forecast Value Added is still tracked by hand, this is also the point to weigh spreadsheet-based tracking against purpose-built planning software, since manual FVA tracking tends to fall behind once a portfolio grows past a few hundred SKUs.
Where Oritiq Fits
This is where an AI-powered planning layer earns its place, sitting over your existing ERP rather than replacing it, including where that ERP runs as a single central instance across multiple plants. Oritiq’s demand planning platform links phased-in and phased-out items so they read as one continuous demand stream instead of a false cold start, and flags supply-constrained periods so the demand series reflects what would have sold rather than what could be shipped. It supports dynamic forecasting that recalculates through the month, at a cadence such as daily or per shift, instead of locking in a single point forecast at the start of the month. Run it against two years of your own history and compare the output to what your current process would have produced.
FAQs
How do you reduce forecast error in supply chain planning?
The short answer to how to reduce forecast error is sequence, not tools. Repair demand history for phase-ins and shortages first, classify each item’s forecastability, forecast at the level the decision needs, and only then adjust the model or bring in a new one.
What is a good forecast accuracy benchmark for organisations?
There is no single number that fits every item. A smooth, high-volume item can support high accuracy. A lumpy, low-volume item cannot, regardless of method. Organisations that want to improve forecast accuracy do better setting benchmarks per demand class rather than applying one target across the whole portfolio.
Can AI reduce forecast error on slow-moving or lumpy items?
AI helps less here than on regular demand. Lumpy items are better served by a buffer and review policy than by a sharper forecast. Statistical methods built specifically for intermittent demand still tend to outperform general machine learning approaches on these items.
Why does forecast accuracy look better at an aggregate level?
Errors that cancel out across items or locations shrink once totals are summed to category or network level. A report can show improving accuracy while every SKU-location decision underneath stays just as wrong. Always measure at the level the decision is made.
What is Forecast Value Added and why does it matter?
Forecast Value Added compares each step, statistical baseline, system output, and human override against the step before it. If an adjustment does not beat the naive forecast it replaced, it is not adding value, whatever confidence the planner has in it.
How do supply shortages distort demand history?
When an item is unavailable, recorded sales fall even where demand did not. A model trained on that period learns that demand was small, when the real constraint was supply. Restoring the demand that would have sold, before forecasting, fixes this at the source.
How long does it take to see lower forecast error?
History repair and forecastability classification usually show results inside one planning cycle, often four to six weeks. Level and override fixes typically take a full S&OP cycle or two to show up in the numbers, since they change what gets measured, not only what gets forecast.