Sales Forecasting & Revenue Analysis
The forecast isn't inaccurate. It's biased.
Prepared for Redbud Fabrication, Inc.
00Summary of findings
Redbud's forecast looks acceptable in aggregate: 5.0% mean absolute error over twenty-four months, with a mild +3.7% tendency to over-forecast. On that evidence the forecast is a minor irritation, not a problem.
The aggregate is misleading, and by a specific factor. Two segments are wrong in opposite directions and largely cancel each other out. Agricultural and construction machinery is over-forecast by 18.7%. Electrical equipment is under-forecast by 20.1%. In dollars, $237,000 a month of over-forecast is netted against $121,000 a month of under-forecast, leaving an aggregate of $116,000 that conceals $358,000 of actual error.
The aggregate understates the problem by 3.1 times.
It is also bias rather than noise, and that distinction is provable rather than rhetorical. For both problem segments, mean absolute error and mean signed error are the same number — every miss lands on the same side. A random forecast produces errors in both directions. This one does not.
Do not hire the demand planner and do not buy the forecasting software. Neither addresses the cause. Separate the number sales commits to from the number operations plans against, publish bias by segment every month, and the accuracy follows. Combined cost of the two purchases under consideration is $155,000 a year; the recommendation costs nothing in software or headcount.
01The question
Redbud misses its forecast most months and has two proposals on the table to fix it: a demand planner at roughly $95,000 loaded, and a forecasting package at about $60,000 a year. Together, $155,000 a year against a problem nobody has yet measured.
How wrong is the forecast, in which direction, where — and would either purchase address the cause?
The order of those clauses matters. "How wrong" is the question Redbud asked. "In which direction" is the one that turns out to matter, and it is not the same question.
02Method
Twenty-four months of forecast submissions were pulled from the CRM at segment level and matched to booked revenue from the ERP. Forecasts were compared at the vintage they were submitted, not as later revised — a forecast quietly corrected in month two is not a forecast that was right.
Two measures are used throughout, and the difference between them is the whole analysis:
- Mean absolute percentage error (MAPE) — how far off, ignoring direction. This is what Redbud reports today.
- Mean percentage error (MPE), or bias — how far off, keeping the sign. Over-forecasts and under-forecasts cancel.
Reporting only MAPE tells you the forecast is imprecise. Reporting both tells you whether it is imprecise or systematically wrong, and those have different causes and different fixes.
03What the aggregate says
| Aggregate measure | Value | Reads as |
|---|---|---|
| Mean absolute error (MAPE) | 5.0% | Respectable for a job shop |
| Bias (MPE) | +3.7% | Mild optimism, easily lived with |
| Months over-forecast | 19 of 24 | The first thing that should worry you |
Two of those three numbers are reassuring. The third is not: a forecast that is honest but imprecise should land above actual roughly half the time. Landing above in nineteen months of twenty-four is not imprecision. It is a thumb on the scale.
04What the aggregate conceals
Breaking the same twenty-four months out by segment:
| Segment | Share | MAPE | Bias | $ per month |
|---|---|---|---|---|
| Ag & construction | 34% | 18.7% | +18.7% | +$201,917 |
| Electrical equipment | 16% | 20.1% | −20.1% | −$104,219 |
| Transportation | 27% | 7.3% | +4.2% | +$35,338 |
| General industrial | 23% | 4.0% | −2.3% | −$16,595 |
| Net — what the aggregate reports | +$116,441 | |||
| Gross — what is actually wrong | $358,069 | |||
Gross segment error runs at 11.2% of monthly revenue while the aggregate reports 3.7%. The two largest errors point in opposite directions and cancel on their way up to the summary. Nobody built it that way on purpose, and nobody looking at the aggregate would ever see it.
Worth naming plainly: the segment Redbud under-forecasts most severely is electrical equipment, which the July Market Analysis identified as the only end market growing on both a one-year and a three-year view. Redbud is systematically under-planning the one market the outside data says is its best opportunity — and then, predictably, finding itself short of material when the orders arrive.
05Bias, not noise
The distinction carries the recommendation, so it is worth proving rather than asserting.
For agricultural and construction machinery, mean absolute error is 18.7% and mean signed error is also 18.7%. For electrical equipment, both are 20.1%. Those pairs being identical is not a coincidence and not a rounding artefact — it is what happens when essentially every monthly miss falls on the same side of zero.
Random error is a modeling problem, and a better model or a dedicated planner can reduce it. Directional error is an incentive problem, and no model fixes it. A forecasting package fed by the same people producing the same directional inputs returns the same bias with more decimal places.
The mechanism is not mysterious and is not anybody behaving badly. Ag and construction is where the commission opportunity is concentrated and where the market has been visibly rebounding, so representatives forecast optimistically — and the forecast doubles as the number they are measured against, which makes optimism rational. Electrical equipment is newer, has no established quota history, and is forecast conservatively because nobody wants to commit to a number they cannot yet defend.
Both behaviors are sensible responses to how the forecast is used. That is the finding: the forecast is being asked to be a commitment and an estimate at the same time, and it cannot be both.
06The horizon problem
| Forecast horizon | MAPE | What it mostly reflects |
|---|---|---|
| 1 month out | 2.2% | Booked backlog — barely a forecast at all |
| 2 months out | 3.6% | Backlog plus near-certain repeat orders |
| 3 months out | 5.0% | Actual forecasting — and what purchasing plans on |
Redbud quotes its forecast accuracy using the one-month number, which is excellent and nearly meaningless: at one month the forecast is largely a restatement of the order book. Material lead times mean purchasing commits on the three-month horizon, where accuracy is less than half as good and where all of the segment bias above is fully present.
This is why the problem shows up as inventory rather than as a reporting complaint. The forecast that gets graded is not the forecast that gets spent against.
07What the bias costs
| Cost | Basis | Per month |
|---|---|---|
| Carrying cost on excess inventory | $1,400,000 of excess stock at 22% a year | $25,667 |
| Obsolete and slow-moving write-downs | $180,000 a year, from over-forecast segments | $15,000 |
| Spot-buy premium on unplanned material | Material bought outside contract when under-forecast | $18,400 |
| Inbound expedite freight | Inbound only | $11,200 |
| Total cost of forecast bias | $70,267 | |
That is $843,200 a year, or 2.2% of revenue.
What is deliberately not counted here
Three costs that a forecasting analysis could plausibly claim have been left out, because the March Constraint Analysis already charged them to the press brake:
- Premium outbound freight — $31,000 a month
- Overtime — $45,800 a month
- Quotes lost on lead time — $58,500 a month
Adding them here would inflate this analysis by $135,300 a month and would double-count against a document Redbud already has. Where two analyses could each claim the same dollar, it belongs to whichever one can actually remove it — and those three go away when the brake constraint does, not when the forecast improves.
The spot-buy premium is the line most likely to grow. Steel rose 11.1% between January and June 2026, so material bought reactively outside contract now carries a materially worse penalty than the two-year average in this table reflects.
08How to challenge this
Assumptions worth arguing about
- That $1,400,000 of inventory is genuinely excess. This is the largest single line and the softest. It is derived from stock turns on over-forecast segments against turns on the rest. Check: a physical review of ag and construction raw stock against the current order book, which takes a day and settles it.
- That segment bias is stable rather than a phase. Twenty-four months is enough to distinguish bias from noise, but the ag rebound is recent. If ag demand keeps accelerating, today's over-forecast becomes tomorrow's accurate forecast without anyone changing anything. Check: re-run the bias split quarterly; the recommendation includes this.
- That better reporting changes behavior. The recommendation assumes publishing bias by segment causes it to shrink. Usually true, because it makes optimism visible and attributable. Not guaranteed — if the commission structure keeps rewarding optimism strongly enough, measurement alone will not overcome it. Check: if bias has not halved within two quarters, the incentive design is the next thing to look at.
- That 60% of the cost is recoverable. Some inventory buffer is deliberate and prudent. Removing all of it would trade one problem for another, so the recommendation targets the excess, not the buffer.
09Recommendation
Separate the commitment from the estimate, and measure bias where it lives. Sales continues to commit to a number it is accountable for. Planning uses a separate figure — the statistical baseline adjusted only for known events — and that is what purchasing spends against. Neither purchase under consideration is needed to do this.
| Weeks | Action | Measure of done |
|---|---|---|
| 1–2 | Split the forecast into a commitment number and a planning number; purchasing plans against the second | Both numbers exist and differ, with no argument about which is "right" |
| 2–4 | Build the statistical baseline from 24 months of actuals, by segment | Baseline reproduces history within the segment MAPEs above |
| 4 | Publish bias by segment monthly, alongside MAPE — never MAPE alone | Every segment's signed error visible to the people producing it |
| 4–8 | Reforecast electrical equipment upward against the market evidence | Electrical bias inside ±8% |
| Quarterly | Re-run the bias split; revisit the software question with real evidence | Gross segment error under 6% of monthly revenue |
The last row is deliberate rather than dismissive. Forecasting software may well be worth buying — but a tool bought to fix a bias problem it cannot fix will be judged a failure and blamed for it. Fix the directional error first, then evaluate the software against what is genuinely left, which will be a smaller and much more honest question.
Expected effect is roughly $506,000 a year — about 60% of the bias cost — against $155,000 a year for the two purchases that would not have addressed the cause.
10Scope and fee
| Engagement | Scope | Fixed price |
|---|---|---|
| Sales Forecasting & Revenue Analysis | Forecast and ERP extraction, bias decomposition by segment and horizon, cost quantification, this document and a working session | $7,500 |
| Implementation support (optional) | Building the split forecast process, the statistical baseline, and the monthly bias reporting, through one full cycle | $16,000 |
| Total if both are taken | $23,500 | |
Ongoing monthly bias reporting is available as part of an analytics arrangement rather than a project — and it is the piece that keeps this from recurring, since bias returns quietly the moment nobody is publishing it.