Sales forecasting management a demand management approach

1. 1Diagnose the credibility problem before choosing a fix. The source says credibility breaks down when users lack confidence in the source of data or in the controls around it. Ask each forecast user (marketing, sales, finance, production/purchasing, logistics) one question: which number do you plan against, and why?
2. 2Map the forecasting process end to end. The source notes that in many companies no one person understands the entire process, and that firms ask for a system before they can describe the process. Write down who produces each forecast, from what data, with what technique, and who receives it. Do this before any tool decision.
3. 3Count the forecasts in circulation. Under the independent approach, each function builds its own forecast for its own needs: finance does corporate-level dollars 1 to 5 years out, production does 6-month SKU forecasts, and logistics does monthly SKU-by-location forecasts. Several unreconciled numbers means there is no single source of truth. Retire them in favor of one common forecast.
4. 4Choose an organizing approach that produces a common number. The source lists four: independent, concentrated, negotiated, consensus. Concentrated bases the forecast on one function's requirements. Negotiated can be biased by power imbalance between departments. Consensus uses a committee with a designated leader, takes input from every department, and builds one forecast from that input (not a blend of departmental forecasts). Consensus is resource intensive in time and people, so adopt it only if you can staff it.
5. 5Build Forecasting C3 into the process. Communication is the shared written, verbal, and electronic information. Coordination is the formal structure and required meetings. Collaboration is common goal-setting, including common forecasting performance goals. Set the meeting cadence and the shared performance goals explicitly.
6. 6Remove islands of analysis. Consolidate the scattered spreadsheet templates ('spreadsheet mania'), the differing data sources, and the unexamined assumptions. Control the source of the data, document the assumptions behind each model, and move toward linked systems with consistent data.
7. 7Publish performance measures with the forecast. Stage 1 companies leave metrics out of their systems and reports. Report percent error by forecasting level and by time horizon, and set a goal for each. Use MAPE, the most common measure (52% of Phase 2 respondents), and weight statistics by volume if you roll up across products.
8. 8Audit subjective inputs. Ask how credible the values from salespeople, channel members, and executives are, and to what degree each group 'game plays' when providing forecasts. Treat the new-product forecast separately: the source warns that poor new-product accuracy hurts the forecasting group's credibility.

The source's own illustration is the independent approach. Finance builds a corporate-level dollar forecast 1 to 5 years out. Production builds 6-month SKU forecasts for scheduling. Logistics builds monthly product/location forecasts for distribution. Each is built for a different purpose, likely from different data, and none is reconciled to the others, so the numbers differ. Layer on a Stage 1 systems picture: forecasters use different spreadsheet templates on different computers, with little control over data sources or assumptions. One Phase 4 executive called this 'spreadsheet mania'. Users then lack confidence in where the data came from and whether controls exist, which the source says breaks credibility. The remedy follows the passages. First, document the actual management process. Next, form a consensus committee with a designated leader and representatives from each function. The committee takes departmental input, including the data and assumptions behind it, and issues one common forecast. Finally, publish percent error by level and horizon against stated goals. Planning functions then work from one number whose error is visible, not from three private ones.

• If users keep maintaining their own forecasts, then you are in the independent approach or have a trust gap. Fix the shared process and data lineage before you buy or change a system.
• If one department dominates the forecast conversation, then the negotiated or concentrated approach is biasing the number. Move toward a consensus committee with a designated leader, if you can resource the time and people.
• If the process is not documented or nobody owns the whole of it, then do not select a system yet. Map the process first, because systems built without a process are a backward approach.
• If subjective inputs come from salespeople, channel members, or executives, then test them for game playing and track their error separately before blending them in.
• If new-product forecast accuracy is poor, then report it separately from established products and use the supplementary procedures the source points to (Chapters 5, 6, 9), so it does not drag down trust in the whole forecast.

• Treating credibility as an accuracy problem only. Accuracy and credibility were each rated important by 92% of Phase 2 respondents, which makes them separate criteria. The tell: accuracy is decent, yet functions still keep their own numbers.
• Buying a system to fix trust. The tell: the team cannot describe the management process, and the system request comes first.
• Calling it consensus when it is only a blend of departmental forecasts. The source says a genuine consensus builds a common forecast from input, not from the departments' individual forecasts. The tell: the final number is an average or a compromise of separate submissions.
• Leaving metrics out of reports. The tell: reports go out with no error statistics, or error is calculated off-line and sits outside the main reports.
• Letting power imbalance or gaming shape the number. The tell: forecasts from one function consistently match that function's own targets, with no visible check on the inputs.

• ▸Forecasts in circulation: list every forecast (owner, level, horizon, purpose). More than one per level and horizon is a credibility risk.
• ▸Process owner: name the person who understands the entire forecasting process, or record 'none'.
• ▸Organizing approach: independent / concentrated / negotiated / consensus. Note any power imbalance among departments.
• ▸C3 status: Communication (information shared), Coordination (required meetings), Collaboration (common performance goals). Rate each as present or absent with evidence.
• ▸Data lineage: for each forecast, list the data source, the spreadsheet or tool used, and the documented assumptions. Flag islands of analysis.
• ▸Performance reporting: measure used (e.g., MAPE), percent error by level and by horizon, goal by level and horizon, and whether the measure is in the report or off-line.
• ▸Subjective input credibility: for salespeople, channel members, and executives, record observed error and any evidence of game playing.
• ▸New-product forecasts: list accuracy separately and the procedures used to supplement them.
• ▸User confidence check: for each function, record which forecast it plans against and why, in its own words.

1. 1Separate the forecast from the plan and the target before measuring anything. The forecast projects demand under stated environmental assumptions, and accuracy is its measure. Plans are judged on meeting forecasted demand effectively and efficiently. Targets are meant to be met or exceeded. Score the forecast against actual demand, never against quota.
2. 2Write down the environmental assumptions behind each forecast: price, promotion, competitor actions, economy. You need them later to explain variances.
3. 3Fix the level of measurement. Name the forecasting hierarchy level, the time horizon and the time interval, for example SKU by location. Report accuracy at each level where a decision is made, because error at one level says little about error at another.
4. 4Compute the actual-error measures for each series. Use Mean Error to see signed bias. Use Mean Absolute Error so over- and under-forecasts do not cancel. Use Sum of Squared Errors or Mean Squared Error when large misses matter most.
5. 5Convert to measures relative to a perfect forecast. Calculate Percent Error per period, then MAPE across periods. Keep a Year-to-Date MAPE for a running view. Percent measures let you compare products of different volume.
6. 6Benchmark against a reference forecasting technique, not only against perfection. Then use the multidimensional and multi-product measures to roll accuracy up across products without hiding the weak ones.
7. 7When error is large, separate two causes. First, check whether the environment differed from the stated assumptions. Second, check whether the technique assumes patterns the history does not match. Simple exponential smoothing assumes only level and noise, so trend or seasonality in the history produces consistent error.
8. 8Set the accuracy you need from the consequences of error at each level, horizon and interval. Weigh the extra accuracy against management, operations and marketing forecasting costs, and against customer satisfaction, as a forecasting ROI decision. Do not assume more accuracy is always better.

The passages give no numeric accuracy case, so this uses hypothetical numbers to show the mechanics; the pattern it shows comes from the text. A product has rising sales. Actuals for four months are 100, 110, 120, 130, and simple exponential smoothing forecasts 100, 100, 110, 120 (hypothetical values). Errors (actual minus forecast) are 0, 10, 10, 10, so Mean Error is +7.5 and Mean Absolute Error is 7.5. Every non-zero error has the same sign, which is the bias signature. The source explains it: simple exponential smoothing assumes only level and noise, so when trend exists it "will consistently err." Percent errors are 0%, 9.1%, 8.3% and 7.7%, so MAPE is about 6.3%. That MAPE looks modest, but Mean Error equal to Mean Absolute Error exposes the one-directional miss. The fix is a technique that models trend, such as exponential smoothing with trend, not an excuse about the environment. Then re-measure at the same hierarchy level and horizon.

• If someone proposes judging the forecast by whether sales hit quota, then separate the two: accuracy belongs to the forecast, and meeting or exceeding belongs to the target.
• If Mean Error is close to Mean Absolute Error (errors persistently one-signed), then suspect the technique omits a pattern such as trend or seasonality, and change the technique before tuning.
• If an accuracy miss is explained by the environment differing from the stated assumptions, then record the explanation to improve future forecasts, not to excuse the miss.
• If accuracy is reported only at an aggregate level, then also measure at the level where the decision is made (e.g., SKU by location) and at the relevant horizon and interval.
• If demanding tighter accuracy, then compare the consequences of error and the added forecasting costs against the benefit. Do not treat accuracy as "the more the better."

• Confusing forecast with target or plan: forecasts get padded or trimmed to motivate people. The tell is that the forecast always equals quota.
• Reporting one accuracy number at one level: the aggregate looks healthy while SKU-by-location errors drive stockouts or excess. The tell is that planners distrust a "good" MAPE.
• Using absolute or percent measures only: persistent over- or under-forecasting stays hidden. The tell is a tolerable MAPE alongside one-signed signed errors.
• Using variance explanations as excuses: the reason is logged but nothing changes in the assumptions or technique. The tell is the same environmental reason appearing every period.
• Chasing accuracy regardless of cost: effort goes into series where error carries little consequence. The tell is rising forecasting cost with no change in customer satisfaction or operating results.

• ▸Series and hierarchy level: name the product/location combination (e.g., SKU by location) being measured.
• ▸Horizon and interval: state the forecast lead time and the period length (e.g., monthly).
• ▸Stated environmental assumptions: list price, promotion, competitive and economic assumptions behind the forecast.
• ▸Mean Error (bias): record the signed average error and note its direction.
• ▸Mean Absolute Error / MSE: record the size of typical error and of large misses.
• ▸MAPE and YTD MAPE: record the percent-error average for the period and for the year to date.
• ▸Reference-technique comparison: record the accuracy of the benchmark technique alongside yours.
• ▸Variance explanation: note how the actual environment differed from the stated assumptions.
• ▸Required accuracy and cost check: state the consequence of error at this level and whether the forecasting cost is justified.

1. 1Name the cost lines before you touch the forecast. The passages give four planning-driven lines: inventory carrying cost (cost of money, obsolescence/shrinkage/insurance risk, facility cost for receiving, handling and shipping), production overtime and expediting, procurement price (spot vs. long-term contract), and transshipment between distribution centers. Add outsourcing to higher-cost alternatives as a fifth line where it applies. Give each line a baseline dollar figure from your own books.
2. 2Trace each cost spike to its demand source. Ask whether the demand hitting you is independent (end-use customers) or derived (a customer's ordering policy). In the ballast case, the 10,000-unit spikes were derived from a contractor's schedule, not random consumer demand. Forecasting derived demand as if it were independent produces the spikes and the expedite costs.
3. 3Where demand is derived, plan it instead of forecasting it. Get the customer's order timing ahead of your fill lead time, using a price incentive if needed. The ballast company's lead time was 3 weeks and it asked for orders 5 or more weeks ahead, so the gap between notice and lead time is what removes expedite, overtime, spot buying and outsourcing.
4. 4Where demand is multi-tier, share point-of-sale data and planned orders up the chain so each tier plans instead of forecasting its customer. In the Figure 1.2 example, system-wide inventory fell 13.81% (5,105 to 4,400 units) and safety stock fell 63.8% (1,105 to 400). Expect the savings to be unevenly distributed (see the decision rules).
5. 5Estimate the improvement in forecast accuracy. Then trace its effect on stockouts (which products, by how much), inventory dollars reduced, production savings, procurement savings and transshipment savings. These are the passage's seven ROI questions; answer all of them, not only the accuracy one.
6. 6Quantify the inventory line in three components. Multiply the inventory dollars removed by the cost of money. Then add the reduced risk costs (obsolescence, shrinkage, insurance) and the reduced facility costs. The sum is the carrying-cost saving.
7. 7Subtract the investment: systems, staffing, training, performance rewards. Management costs are the 'investment' and reduced operations costs are the 'return'. Present the result as a proposal in the format of Figure 2.5, ending in profit impact, not accuracy percentage.
8. 8Document only what you can defend. When the revenue effect cannot be fully measured, count it only where it is observable. The Figure 2.5 company counted a sale only when inventory was available rather than a stockout. Keep company-specific metrics for cost and service, since the passages note these cannot be standardized the way accuracy can.

The Figure 2.5 case is the model. A company improved forecast accuracy by product and by distribution center and documented the result. Safety stock fell by $45,000,000 of inventory, which cut inventory carrying cost (cost of money, risk costs, facility costs) by $5,000,000 per year. Better forecasts let it buy more raw materials on long-term contracts instead of the spot market, saving $1,000,000 annually. It also needed fewer transshipments from the East Coast to the West Coast distribution center, after shipping product from the Midwest plant to the wrong coast. The source reports operating costs down $7,000,000, increased sales of $2,000,000 from better in-stock positions, and increased profit of $7,010,000. The passage text supplied is cut off before the transshipment dollar figure, so do not back-solve it. The ballast company shows the production side. Its 10,000-unit spikes arrived one week before need against a 3-week fill time, forcing expedited production, higher inventory, spot-market procurement and outsourcing. It offered a 3% discount on orders over 10,000 units placed 5 or more weeks ahead. On units costing more than $20, that is a discount above $6,000 per order, and the company gets 5 weeks' notice for a 3-week job.

• If a cost spike comes from a customer's ordering pattern (derived demand), then build an advance-order incentive rather than a better spike forecast. The ballast discount did exactly this.
• If your notice window is shorter than your fill lead time, then expect expedite, overtime, spot-market and outsourcing costs. Set the incentive so notice exceeds lead time (5 weeks vs. 3 weeks).
• If a supply chain partner would carry most of the savings but not the cost, as with upstream firms that gain the most safety-stock reduction and retailers who gain none, then design rewards so partners are fairly compensated. Cooperation will not follow the economics on its own.
• If a benefit cannot be fully measured (such as the revenue effect of improved service), then count only the documented portion, for example sales made because stock was available. Do not claim the rest.
• If you cannot answer the cost questions (stockouts, inventory dollars, production, procurement, transshipment savings), then do not present the proposal. Accuracy gains alone do not reach C-level decision makers.

• Reporting accuracy gains as the result. The tell: the proposal stops at an accuracy percentage, with no link to inventory, overtime, procurement or transshipment dollars.
• Forecasting derived demand as if it were random. The tell: large irregular spikes, a flat forecast, and a recurring scramble of expedites and outsourcing whenever a contractor's big order arrives.
• Counting inventory savings as a single number. The tell: the inventory dollars reduced are quoted without cost of money, risk costs and facility costs, so the carrying-cost saving cannot be checked.
• Ignoring partner incentives. The tell: a POS-sharing or demand-planning initiative stalls at the retailer or customer, who sees no inventory reduction for itself.
• Leaving out the investment side. The tell: savings are stated with no figure for systems, staffing, training or performance rewards, so no ROI can be computed.

• ▸Demand source and type: Note whether the spiking demand is independent or derived, and from whose ordering policy.
• ▸Notice vs. lead time: Record the customer's current notice (weeks) against your fill lead time (weeks).
• ▸Inventory reduction: Enter total inventory dollars expected to be removed (safety stock) and the cost of money.
• ▸Risk and facility cost reduction: Enter the savings from obsolescence, shrinkage, insurance, and from receiving, handling and shipping.
• ▸Production savings: Enter the expected reduction in overtime, expedited operations and outsourcing premiums.
• ▸Procurement savings: Enter the price difference between long-term contract and spot buying, times volume moved.
• ▸Transshipment savings: Enter the cost of moves between distribution centers avoided.
• ▸Documented revenue effect: Enter only sales you can show were made because stock was available.
• ▸Total investment: Enter systems, staffing, training and performance rewards, then compute net return.

1. 1Locate your position in the chain. Only the company that serves the end-use customer sees independent demand. Everyone upstream sees derived or dependent demand. Write down which type of demand hits you, because the techniques, systems and processes differ for each, and the mismatch is what hurts customer service levels.
2. 2Define satisfaction in the terms the construct uses: product availability, on-time delivery, and fulfillment of demand. Tie each to a count you already hold (orders shipped complete, orders shipped on the promised date, orders you could not fill). The passages give no scoring formula, so use your own records.
3. 3Find the demand events that break service. Look for large orders that arrive with less notice than your fill time. The ballast company needed 3 weeks to fill 10,000 units but was told one week before. List every customer or order type with that gap between notice and lead time.
4. 4Price the failure. For each gap event, list the costs the passages name: expedited production, higher inventory, spot-market procurement, outsourcing to higher-cost alternatives, and thinner margin on the order. These are the cost of late or short fulfillment.
5. 5Replace forecasting of the spike with planning of the spike. If the spike is derived demand (it comes from a customer's ordering policy), get the customer's order timing instead. Offer an incentive tied to lead time: order early and share in the savings.
6. 6Set the incentive notice window longer than your fill time. The ballast company set 5+ weeks of notice against a 3-week fill, which leaves 2 weeks to produce normally with no overtime or outsourcing. Check that the discount is smaller than the avoided expediting cost.
7. 7Pass real demand signals up the chain. Where you can, share point-of-sale (POS) demand and planned orders with upstream partners instead of letting each one forecast the next one's orders. This removes the stacked safety stock that inflates upstream demand.
8. 8Align measurement and rewards across partners. Customer satisfaction and cost reduction are the joint measures of supply chain success. Partners need rewards matched to those measures or they will not collaborate.

An electrical-products company supplies ballasts for office-building retrofits. A contractor needs 10,000 units, but tells the company the week before they are needed, and the company takes 3 weeks to fill an order that size. The 2-week shortfall forces expedited production, higher inventory, spot-market procurement and outsourcing, so the order earns little or no money and delivery is at risk. The company recognizes that this is derived demand, set by the contractor's ordering policy, and stops trying to forecast the spikes. It offers contractors a 3% discount on orders above 10,000 units placed five or more weeks before they are needed. At more than $20 a unit, 3% of a 10,000-unit order is over $6,000 for the contractor. The company now knows about a 3-week-fill spike 5 weeks ahead and can produce it in normal time. The passage reports millions in annual savings and increased sales and market share with contractors. Satisfaction comes from the contractor getting a price break and an order filled on time, while the company stops absorbing expediting costs.

• If a customer's orders arrive with less notice than your fill time, then treat the demand as derived and plan it with the customer, not forecast it.
• If an incentive costs less than the expediting, overtime, spot-buy and outsourcing it avoids, then offer it, and set the notice window longer than your lead time.
• If you sit upstream of the end-use customer, then ask for POS data and planned orders rather than forecasting the next company's orders.
• If every partner adds its own safety stock for +/-10% error, then expect the error to compound. The source example reaches 1,105 units of safety stock on 1,000 units of true demand.
• If a partner would lose from collaborating, then fix measurements and rewards first, or the planning arrangement will not hold.

• Forecasting derived demand as if it were independent demand. The tell is repeated large spikes you did not see coming and constant expediting.
• Each tier adding its own safety stock. The tell is inventory rising at every step upstream, as in 1,100, 1,210, 1,331 and 1,464 units, while end-use demand is 1,000.
• Offering a discount with a notice window shorter than your fill time. The tell is early orders that still force overtime or outsourcing.
• Treating satisfaction as a survey score with no link to availability, on-time delivery and fill. The tell is a satisfaction number that moves independently of shipment records.
• Pushing planning on partners without rewards for them. The tell is partners that keep forecasting on their own and ignore the shared demand data.

• ▸Customer / order type: name the account or order pattern that causes service problems.
• ▸Demand type at your position: independent, derived or dependent, and why.
• ▸Typical order size: units per order, plus unit cost or price.
• ▸Fill lead time (weeks): how long it takes to fill an order of that size normally.
• ▸Notice received (weeks): how far ahead the customer actually tells you.
• ▸Gap (weeks): lead time minus notice; a positive number means expediting.
• ▸Gap costs incurred: list expedited production, extra inventory, spot buys, outsourcing and margin lost.
• ▸Proposed incentive and notice window: the discount % and the minimum weeks of notice (longer than fill lead time).
• ▸Service result to track: orders filled on time and complete, before and after the incentive.

1. 1Start with the P&L, not the forecast. Take sales revenue and subtract ALL costs, fixed and variable. This is not gross margin. The result is the profit of the business unit. In the source's company this is $2,000,000,000 of sales less $1,900,000,000 of total costs, which leaves about $100 million of annual profit.
2. 2Define the capital base: working capital (primarily accounts receivable and inventory) plus fixed capital. Leave retained earnings out. They are a board and shareholder financing decision, and you are isolating the effect of operations decisions.
3. 3Compute the baseline: return on shareholder value = profit ÷ capital invested. It is the percent of invested capital returned as profit each year. Record it before changing anything.
4. 4Translate better forecast accuracy into two operational channels. (a) Better operational plans and execution lower costs per dollar of sales. (b) Better customer service raises sales. Accuracy gains that reach neither channel contribute nothing to shareholder value.
5. 5Quantify the cost channel line by line. Estimate production savings (overtime, expedited operations), procurement savings, and transshipment savings. Estimate the reduction in risk costs (obsolescence, shrinkage, insurance) and in facility costs (receiving, handling, shipping) that follows from lower inventory.
6. 6Quantify the capital channel. Estimate how far inventory, the working-capital component, falls when forecasts improve. This lowers the denominator.
7. 7Quantify the revenue channel. Link better availability and order condition (accurate picking, little damage) to retained or added sales. State these assumptions explicitly.
8. 8Recompute profit and capital, then the new return on shareholder value. Present both versions side by side to C-level executives. Stating accuracy improvement alone is not enough, because executives need the final shareholder-value impact to make an informed decision.

The source's actual example is a company with $2,000,000,000 in sales revenue and $1,900,000,000 in total costs, so annual profit is about $100 million. (The source says its numbers were slightly altered to protect the company.) The source excerpt cuts off before the capital and improvement figures, so none are invented here. The method is as follows. Profit divided by capital invested gives the baseline return on shareholder value. A more accurate forecast improves execution, which lowers total costs per dollar of sales. Better service lifts sales revenue. Both changes raise the numerator, profit. Lower inventory shrinks working capital, which lowers the denominator. Either move raises the ratio, and together they compound. The proposal you build, like the source's Figure 2.5, lays the before and after profit and capital side by side. It states the result as a change in return on shareholder value, not as a change in forecast accuracy.

• If a forecasting improvement proposal reports only accuracy gains (for example, a lower MAPE), then do not send it to executives. Convert it to profit and capital effects first.
• If a benefit cannot be traced to lower costs per dollar of sales or to increased sales through service, then exclude it from the shareholder-value case.
• If inventory falls, then count both effects: the smaller capital base and the lower obsolescence, shrinkage, insurance, and handling costs.
• If you are tempted to include retained earnings, then drop them. Only operations-driven capital (receivables, inventory, fixed assets) belongs in the model.
• If costs are being measured, then use all costs, not just cost of goods sold. A gross-margin view hides the savings in production, procurement, and transshipment.

• Stopping at accuracy. The tell: the business case cites percent error but contains no profit or capital figures. The source notes no company succeeded from accuracy alone.
• Using gross margin instead of total cost. The tell: the profit line subtracts only cost of goods sold, so overtime, expediting, and facility savings never appear.
• Counting only the profit side. The tell: inventory reduction is mentioned as a cost saving, but the capital base in the denominator is unchanged.
• Treating a percent error as self-explanatory. The tell: a '2% error' is called good without volume or value. 2% of 100 low-value units is trivial, while 2% of 1,000,000 units a month is 20,000 units of a product vital to customers.
• Mixing in retained earnings or financing choices. The tell: the capital base changes with dividend or board decisions, not with operations.

• ▸Baseline sales revenue: annual revenue before the improvement.
• ▸Baseline total costs (fixed + variable): all costs, not just COGS.
• ▸Baseline profit: revenue minus total costs.
• ▸Baseline capital invested: working capital (receivables, inventory) plus fixed capital; exclude retained earnings.
• ▸Baseline return on shareholder value: profit ÷ capital.
• ▸Cost savings estimate: production (overtime, expediting), procurement, transshipment, plus inventory-related risk costs (obsolescence, shrinkage, insurance) and facility costs (receiving, handling, shipping).
• ▸Inventory/capital reduction: expected fall in working capital from the better forecast.
• ▸Sales increase from service: expected added or retained revenue from better availability and order condition, with its assumption.
• ▸Projected profit and capital: baseline figures adjusted for all of the above.
• ▸Projected return on shareholder value and change vs. baseline: the figure to present to C-level executives.

1. 1Define the three C3 components separately before scoring anything. Communication = written, verbal, and electronic information shared between functions. Coordination = formal structure and required meetings between two or more functions. Collaboration = orientation toward common goals (common forecasting performance goals), including with key customers.
2. 2Map who forecasts today. For each function (finance, marketing, sales, production, logistics) record the format it uses: finance's corporate dollar forecast 1 to 5 years out, production's 6-month SKU forecast, logistics' monthly SKUL forecast. Differing formats are not the problem; missing integration is.
3. 3Classify your approach: independent (little or no communication, no coordination, no collaboration), concentrated (one department owns the forecast), negotiated (departments bargain from their own forecasts), or consensus (a committee builds one common forecast). Name the owning department if concentrated.
4. 4Test communication gaps with a concrete question: does production or logistics see marketing's promotional schedule before forecasting? List each input one function needs from another and mark whether it arrives, in what format, and when.
5. 5Test coordination: list the formal meetings and the distribution structure for the official forecast. If none exist, or all departments receive it only as printed reports that must be re-keyed, the coordination and systems layers are weak.
6. 6Test collaboration: check whether the official forecast carries its owner's bias. Marketing ownership tends to yield capacity-unconstrained forecasts; operations ownership tends to yield capacity-constrained forecasts that ignore market demand. Also check whether a power imbalance dominates negotiation.
7. 7If you move toward consensus, form a committee with one representative per function and a designated head. The committee builds a common forecast from each department's informational input, not from averaging departmental forecasts. Budget time and personnel, because the approach is resource intensive.
8. 8Document the management process before choosing a system. Identify who understands the whole forecasting process, then match the system (common data warehouse, EDI or web-enabled interfaces with customers and suppliers) to that process, not the reverse.

The source describes the independent approach in a typical company. Finance builds a corporate dollar forecast 1 to 5 years out, production builds a 6-month SKU forecast for scheduling, and logistics builds a monthly SKUL forecast for distribution. None of them talk. Production and logistics cannot see marketing's promotional schedule, so their forecasts miss known demand events. The departmental forecasts also cannot feed corporate planning. Moving to the concentrated approach with logistics as owner adds communication and a distribution structure, but the forecast now comes out at the SKUL level with logistics' horizon and update cycle, and it reflects capacity constraints more than market demand. If marketing owns it instead, the forecast comes out in brand or product-line dollars, on a yearly horizon, and ignores capacity. The fix the source points to is a consensus committee with one member per function and a designated leader. The committee builds one common forecast from each function's input, and shared performance goals reduce the political bias seen in negotiated forecasts. The cost is time and people.

• If departments each build their own forecast and none sees the others' inputs (for example, marketing's promotions), then you are independent. Start with communication, because accuracy is being hurt by missing perspectives.
• If one department owns the forecast and the others only receive it, then expect that department's bias (capacity-unconstrained from marketing, demand-blind from operations). Add input from the other functions and shared goals.
• If departments each submit forecasts and bargain, and one function dominates, then forecasts reflect politics. Move to a common forecast and build commitment to common goals.
• If you have the time and staff to sustain a cross-functional committee, then choose consensus. If you do not, then do not claim consensus; an under-resourced committee gives only the appearance of integration.
• If someone asks which forecasting system to buy, then first ask for the management process that produces the forecast. If nobody can describe it end to end, fix the process before the system.

• Mistaking different formats for lack of integration. The tell: effort goes into forcing one forecast format on all departments, while departments still do not communicate or share goals.
• Calling a concentrated forecast collaborative because it is distributed to everyone. The tell: other departments receive the forecast but their input does not change it, and it shows the owner's horizon, level, and update interval.
• Running a consensus committee that averages departmental forecasts. The tell: the committee works from each department's own number rather than informational input toward a single common forecast.
• Letting a dominant department win negotiations. The tell: the final forecast consistently matches one function's requirements, and other functions treat it as that function's number.
• Buying a system before understanding the process. The tell: no single person can describe the whole forecasting process, and data moves between functions as printed reports that must be manually re-entered, with few outside IT understanding the systems.

• ▸Function and forecast format: for each function, record the level, horizon, and update interval it forecasts (e.g., SKU, 6 months).
• ▸Current management approach: mark independent, concentrated, negotiated, or consensus, and name the owning department if concentrated.
• ▸Communication evidence: list the written, verbal, and electronic information shared between functions, and note what is missing (e.g., promotional schedule).
• ▸Coordination evidence: list formal meetings and the structure for distributing the official forecast.
• ▸Collaboration evidence: record whether common forecasting performance goals exist across functions and with key customers or suppliers.
• ▸Bias check: state whether the forecast leans capacity-unconstrained (marketing) or demand-blind (operations), and which function dominates.
• ▸Systems link: note whether functions share a common data warehouse or exchange data via EDI or web interfaces, or rely on printed reports and re-keying.
• ▸Process owner: name who understands the entire forecasting process, or record 'none'.
• ▸Resource check: list the time and personnel available for a consensus committee, and name its designated head.

1. 1Write down the definitions before you build anything. Keep three separate columns or documents: the forecast (your best estimate of demand), the plan (what the company will do, including supply), and the target (what the company wants). The source's Key 2 is 'Forecast demand, plan supply.' Never let a plan or target overwrite the forecast number.
2. 2Specify what the history measures. Sales, shipments, and demand are three different data series (Chapter 1 lists them as a data-characteristics question). Record which one each forecast is built on. If your history is shipments, flag it as a proxy for demand, and note any period where supply constraints could have distorted it.
3. 3Answer the nine Sales Forecasting Management Questions for your company: customer base narrow or broad; data characteristics; number of forecasts (horizons, intervals, products, channels, locations); new products; regional differences; seasonality; personnel and systems sophistication; budget; accuracy needed. These answers set the approach, so write them down first.
4. 4Segment by customer base. If the base is narrow, build a qualitative forecast from direct customer information. If it is broad (packaged goods to consumers), rely on statistical techniques and judgment adjustments.
5. 5Segment products and locations by the level at which each decision is made. Assign SKU or SKU-by-location forecasts to fixed-model time series (FMTS) techniques that managers select. Reserve subjective executive judgment for aggregates such as product lines, brands, or the corporate level. Do not give executives item-level work.
6. 6Split the work into a reconciled top-down/bottom-up structure. Statistical techniques produce the bottom-up item and location forecasts. A cross-functional committee meets at a set rhythm (for example, quarterly for the long horizon and monthly for adjustments) and applies adjustments at the aggregate level. Reconcile the two so that item forecasts sum to the adjusted aggregate.
7. 7Match technique to the data. FMTS and regression techniques generally ignore seasonality. For highly seasonal products, either do not use them or pair them with a technique that handles seasonality. Forecast regionally different markets separately (for example, U.S. versus Canadian agricultural chemicals).
8. 8Check the approach against your people and systems. If personnel lack statistics or industry experience, train them and limit technique sophistication until they are trained. If systems are 'islands of analysis,' expect the benefits of integrated forecasting to be limited.

An auto parts manufacturer in the source separates the forecasting levels explicitly. Its forecast committee meets once a quarter to subjectively generate a three-year forecast. Once a month it subjectively adjusts product forecasts by product line; for instance, every product forecast in a line may be raised by 3%. Individual product forecasts by inventory location are left to an appropriate FMTS technique chosen by the forecast managers. The authors note that having the committee forecast individual products would waste valuable executive time. This is a reconciled top-down/bottom-up design with segmentation by level. A second case shows customer segmentation: a company selling only to North American automobile assembly plants (annual sales over $50 million) has just 56 customers. Each month it calls every plant's production scheduling department, receives the car production schedule by EDI, and derives a very accurate qualitative forecast from it. Both companies match method to the structure of their business rather than applying one technique everywhere.

• If your customer base is narrow (a few dozen customers), then build the forecast from direct customer schedules and information. If it is broad, then lean on statistical techniques plus judgment adjustments.
• If the decision is at the item or item-by-location level, then use an FMTS technique and keep executives out of it. If it is at the product-line or corporate level, then use committee judgment, applied at a set cadence.
• If the product is highly seasonal, then do not use plain FMTS or regression alone. Pair them with a technique that models seasonality.
• If the market behaves differently by region (for example, a shorter growing season), then forecast each region separately rather than scaling a combined forecast.
• If forecasting staff are weak in statistics or industry knowledge, then train them and cap technique sophistication until the training is done.

• The independent approach: each function builds its own forecast (finance a 1-5 year corporate dollar forecast, production 6-month SKU forecasts, logistics monthly SKUL forecasts). The tell is several different numbers for the same demand, and no communication, coordination, or collaboration among the functions.
• Forecasting shipments and calling it demand. The tell is that the history comes from shipment records and nobody can say whether constrained periods were adjusted.
• Using one technique for everything, such as an FMTS technique on a highly seasonal product. The tell is a forecast that consistently misses the same seasonal months.
• Executives forecasting at the item level, or analysts forecasting only at an aggregate level. The tell is committee time spent on individual SKUs while the aggregate adjustments are not made.
• Forecasting concentrated in one function, for example marketing producing only brand or product-line dollars on a yearly horizon with monthly or quarterly updates. The tell is that downstream functions cannot use the forecast in their own formats, and capacity problems follow (the passage is cut off after 'capacity unc...').

• ▸Forecast / Plan / Target: write the three numbers for the next period in separate cells, and name the owner of each.
• ▸Data series used: sales, shipments, or demand? Note any constraint or distortion in the history.
• ▸Customer base: narrow or broad, with the approximate number of customers.
• ▸Seasonality and regional differences: list the products that are highly seasonal and the regions that must be forecast separately.
• ▸Top-down level: the aggregate being forecast (product line, brand, or corporate), the horizon, the meeting cadence, and the participants.
• ▸Bottom-up level: the item or item-by-location forecasts, the technique assigned to each, and who selects it.
• ▸Reconciliation rule: how aggregate adjustments (such as +3% on a product line) flow down to item forecasts.
• ▸Segment-to-method map: for each product or customer segment, state the method (customer-schedule qualitative, FMTS, regression, committee adjustment).
• ▸Capability gaps: personnel training needs, system interconnections missing, and the budget status.

1. 1Map every system that touches the forecast: where data is stored, where forecasts are developed and analyzed, and where they are distributed. Cover hardware, software, storage, retrieval and transfer. The source defines the construct as all of this, so an incomplete map hides the islands.
2. 2Mark each hand-off between functions as electronic or manual. A hand-off is manual if it moves by printed report and is re-keyed into the receiving function's system. Each manual hand-off is an island of analysis. Note where the format is incomplete and the receiver must add information from other sources.
3. 3Test understanding, not just connectivity. Ask users in each functional area whether they understand what the system does. At the low end of the scale only IT understands it. At the high end the forecast developers and users in every function do.
4. 4Move toward a common data warehouse that users in each function access through their own system-user interface. Replace printed-report transfers first. The source describes this as the sophisticated end of the scale.
5. 5Add data exchange with key customers and suppliers, scaled to need. The range runs from simple EDI to web-enabled interfaces. Start where a partner already holds forecast-relevant information, such as a customer's production schedule.
6. 6Build a suite of tools rather than one technique. Include several time series techniques, and use time series, regression and qualitative methods together (the source's 'and, not or'). Let the system recommend which technique fits each series, and tell the system which forecasts matter most so effort goes there.
7. 7Fit the tool to the problem and the people. Keep complexity under the hood so the interface does not look complex. If personnel lack statistical or industry knowledge, train them and limit technique sophistication until the training is done.
8. 8Tie the system to the whole process, from data input through analysis to reporting and distribution. Check that the forecast is communicated through the system, not around it. Then fund it, since few companies in the studies felt their forecasting budget was adequate.

Two companies from the source show a system fitted to the problem. The first sells a product only to North American automobile assembly plants. It has more than $50 million in annual sales but just 56 customers. Each month its forecasting function calls the production scheduling department of each plant and asks for the car production schedule, which arrives by EDI. A very accurate qualitative forecast is derived from that schedule. The system here is a customer data exchange feeding the forecast directly. Nothing is printed and re-keyed, and no complex statistics are needed. The second company, an auto parts manufacturer, runs a forecast committee. It meets quarterly to set a three-year forecast subjectively. It meets monthly to adjust forecasts by product line, for example raising all products in one line by 3%. Product forecasts by inventory location go to an appropriate fixed-model time series technique chosen by the forecast managers. The committee does not forecast individual products, which would waste executive time. Taken together, the examples show the system routing each forecast to the right tool and the right person.

• If information moves between functions by printed report and is re-keyed, then treat it as an island of analysis and replace it with an electronic link to a shared data source.
• If the customer base is narrow and customers can send schedules electronically, then build the system around direct customer data (EDI or web interfaces) rather than heavy statistical modeling.
• If forecast personnel lack statistical or industry background, then train them and cap technique sophistication until training is done.
• If you are choosing between time series, regression and qualitative methods, then do not choose. Provide all three in the suite and let the system indicate which technique fits each series.
• If a forecast is high-stakes or high-value, then tell the system so it gets attention. Leave low-level detail, such as products by location, to automatic technique selection rather than senior review.

• Buying a sophisticated tool without connecting it. The tell is that users still re-key data from printed reports, so the system's benefits are never realized.
• Systems only IT understands. The tell is that forecasters and users in other functions cannot explain what the system does or why a number changed.
• A single-technique system. The tell is that seasonal products are forecast with fixed-model or regression techniques that ignore seasonality, with no alternative offered.
• Executives forecasting at the wrong level. The tell is senior committees debating individual product forecasts that a technique could produce, while product-line adjustments get no attention.
• Underfunding. The tell is that training and integration gaps persist year after year because no budget is committed, a pattern few companies in the studies escaped.

• ▸Systems inventory: list every system used to store, develop, analyze and distribute the forecast.
• ▸Hand-off list: for each function-to-function transfer, record whether it is electronic or by printed report with re-keying.
• ▸Island count: total the manual hand-offs and name the worst one (incomplete format, added data from other sources).
• ▸Data source: note whether a common data warehouse exists and which functions access it.
• ▸User understanding: record for each function whether its users understand what the system does, or only IT does.
• ▸External exchange: list customers and suppliers connected, and the method (EDI, web-enabled, none).
• ▸Technique suite: record which time series, regression and qualitative tools are available and whether the system recommends techniques.
• ▸Priority forecasts: list the forecasts the system has been told are important and the ones left to automatic selection.
• ▸Gaps and budget: note the training needs, the system fixes, and the committed budget for each.

1. 1Build the scorecard on more than one dimension. The source's performance-measurement chapter covers four: accuracy, forecasting time horizon, costs (management, operations, marketing), and customer satisfaction. A MAPE-only scorecard fails the construct.
2. 2Measure accuracy in three families, each answering a different question. (1) Actual measures: mean error, mean absolute error, sum of squared errors, mean squared error. (2) Measures relative to a perfect forecast: percent error, MAPE, year-to-date MAPE. (3) Measures relative to a perfect forecasting technique, which compare your method against a reference. Report at least one from each family.
3. 3Check bias and stability, not only fit. For regression-based forecasts, R² and adjusted R² are not enough. Use Mallows' Cp: a model is preferred when Cp is approximately equal to its number of predictors k, and Cp well above k signals bias (underfitting) or instability (too many predictors).
4. 4Roll accuracy up across multiple products with a defined multi-product method, since the source lists this as its own topic. Always record the level, horizon, and interval at which each number was measured. A weekly plant-level forecast and a monthly strategic forecast need different horizons and update intervals and cannot be compared on one number.
5. 5Account for forecasting costs in all three buckets: management, operations, and marketing. Compare them with the accuracy gained. The source warns that the costs are not always justified, so end with a forecasting ROI decision: is the next increment of accuracy worth the next increment of cost?
6. 6Add customer satisfaction as a measured outcome of forecast quality, alongside cost reduction. The source defines supply-chain success as cost reduction plus increased customer satisfaction.
7. 7Communicate results to every function that uses the forecast (marketing, sales, finance/accounting, production/purchasing, logistics). Each has different needs, so show each the measures that matter to it.
8. 8Tie the measures formally to evaluation and rewards for the people and partners who shape the forecast. The source says to match performance with measurements and rewards so that all parties are fairly rewarded for overall supply success. Make accuracy and its cost and service effects part of the review, not an after-the-fact report.

The source's regression example shows why one measure is not enough. Model (a) is y = 153.39 + 5.29x1 + 0.98x2 − 3.25x3 with Cp = 2.92 and k = 3. Model (b) is y = 293.31 + 1.30x1 + 2.84x2 − 5.88x3 + 37.53x4 with Cp = 8.16 and k = 4. A scorecard that stopped at variance explained could not separate them reliably. Cp does: model (a)'s Cp sits close to its k, while model (b)'s is far above its k. That gap points to bias or instability, so model (a) is preferred. The same logic applies to human forecasts. In one interviewed manufacturer, the consensus committee was a formality, and marketing and finance had already set the forecast. A scorecard that reports error by source (the quantitative forecast, the committee-adjusted forecast, the finance-influenced number) would show whether the adjustments helped. The source says powerful members, often finance, pull forecasts toward the business-plan revenue figure. Tying rewards to accuracy rather than to agreement with the plan removes the incentive to do that.

• If you report only one accuracy number, add a measure from another family (an actual error, a percent error, and a comparison against a reference technique) before using it to judge anyone.
• If a regression model's Cp is well above its number of predictors, treat it as biased or unstable and do not accept it on R² alone.
• If more accuracy would cost more in management, operations, or marketing effort than it returns, stop investing. The source says the costs are not always justified.
• If forecasts are moving toward a plan number rather than toward demand, add accuracy to the evaluation of the people who adjust them and score the adjusted forecast against the unadjusted one.
• If supply-chain partners must collaborate on the forecast, give them measures and rewards aligned to shared cost reduction and customer satisfaction. Otherwise they have no incentive to share accurate information.

• Single-metric measurement. The tell: the only number reported is MAPE, or R² for models, and nobody can say what the forecast costs or whether customers were served better.
• Measuring at one level and applying the result to another. The tell: a product-level or monthly error figure is quoted to justify a plant-level weekly forecast.
• Measurement with no consequence. The tell: accuracy is reported but not part of reviews or rewards, so the committee's output keeps matching the revenue plan regardless of results.
• Rewards that conflict with the data you need. The tell: distributors deliberately under-forecast new products to avoid inventory liability, and their incentives still reward that caution.
• Chasing accuracy past its payoff. The tell: cost lines (analyst time, systems, marketing effort) grow while the service and inventory benefits are never quantified, and no ROI decision was made.

• ▸Forecast measured (product, level, horizon, update interval): state exactly what is being scored so numbers are comparable.
• ▸Accuracy – actual measures: record mean error, mean absolute error, and MSE or SSE for the period.
• ▸Accuracy – relative to a perfect forecast: record percent error, MAPE, and year-to-date MAPE.
• ▸Accuracy – relative to a reference technique: name the comparison technique and record whether the current method beats it.
• ▸Bias/stability check (models): record Cp and k; flag the model if Cp is well above k.
• ▸Costs: list management, operations, and marketing costs of producing the forecast.
• ▸Customer satisfaction and cost outcomes: record service and cost results that the forecast influenced.
• ▸ROI judgment: state whether the extra accuracy justifies the extra cost, and why.
• ▸Communication and reward link: list who received the results, and which evaluation or reward decision this scorecard feeds.

1. 1Map the process as five stages: data gathering, quantitative analysis, qualitative adjustment, reconciliation, dissemination. Name an owner and an input/output for each. Where nobody owns a stage, the process is already weak there.
2. 2Name one forecasting champion who is the central focus of the function. That person needs the experience and training to make informed decisions on every aspect of managing it. The source finds this common to effective forecasting across many companies.
3. 3Fix the metric before the process. Score every forecast on all three dimensions the source requires: accuracy, cost (management, operations, marketing), and customer satisfaction. Accuracy alone cannot show whether an improvement pays for itself.
4. 4Let the system do the routine work. Run the time series, regression, and qualitative inputs through the system, and have it recommend techniques. Tell it which forecasts matter most. Forecasters then examine only the products where the system has not reached the accuracy management considers acceptable, a threshold set from an ROI analysis of management costs, operations costs, marketing costs, and desired customer service levels.
5. 5Do not set model parameters by habit. For any technique with smoothing constants (α, β, γ), document why each value was chosen and test sensitivity. Prefer the simplest technique that performs equally well; the source says many sophisticated techniques need more data and give no better results.
6. 6Put the qualitative adjustment where the knowledge is. Train adjusters to make judgment forecasts systematically and to document their reasoning, so others can follow and reuse it. The chemical manufacturer decentralized adjustment to the sales force and runs with one forecaster.
7. 7Train by role: people maintaining quantitative forecasts need to know where time series and regression work and fail. Adjusters need to know how to record judgment. Users need to know how the forecasts were built and what their limits are. Treat this as an ongoing cost.
8. 8Audit the whole chain periodically. The source's Phase 4 work covered more than 25 company audits and treats an audit as the starting point for process improvement. Use it to find where the process breaks, not only where accuracy is poor.

A major home appliance company's refrigerator forecast rested on one technique, exponential smoothing with trend and seasonality, which needs three smoothing constants between 0 and 1. The analyst had set α = 0.1, β = 0.2, γ = 0.3. The authors expected a sensitivity analysis behind those values. Instead, he had picked them because every article he read listed the constants as α, β, γ in that order. The data and the model were adequate, but the analysis stage had no tested basis and no documented rationale, so the whole forecast was unreliable. Applying the steps above: assign the champion to review the parameters, test alternative constant combinations and record why one set wins, and check the result against the simpler techniques the source says often perform as well. Then score the outcome on accuracy, cost, and customer service together, not on one number.

• If the system's forecast for a product meets the ROI-derived accuracy threshold, then leave it alone and spend forecaster time on the products that miss it.
• If a simpler technique performs as well as a sophisticated one, then use the simpler one; extra complexity adds data needs and cost with no accuracy gain.
• If the people who know customers and products are in the sales force and the forecasting staff is small, then decentralize qualitative adjustment to them, but require documentation of each adjustment.
• If demand spikes are really derived from a partner's ordering policy, then plan demand with incentives rather than forecast the spikes. The ballast company offered contractors a 3% discount on orders over 10,000 units placed five or more weeks ahead.
• If accuracy gains cannot be tied to cost and customer satisfaction, then do not fund them; require the ROI view first.

• Arbitrary parameters: smoothing constants chosen by convention or habit, as with 0.1/0.2/0.3 from the α, β, γ order. The tell is that nobody can state why a value was chosen or show a sensitivity test.
• No forecasting champion: no one is the central focus, so stages are owned by silos. The tell is that no single person can describe the full flow from data to dissemination.
• Undocumented judgment: qualitative adjustments are made but the reasoning is not recorded. The tell is that others cannot understand or reuse an adjustment, and nobody can say whether it helped.
• Accuracy-only measurement: management cannot judge cost efficiency or customer satisfaction. The tell is that accuracy reports exist but no cost or service numbers sit beside them.
• Forecasting what should be planned: the company chases spikes in what is really derived demand. The tell is expedited production, higher inventory and spot buys, as in the ballast case where the 10,000-unit order arrived one week before a 3-week fill time.

• ▸Stage owner map: for data gathering, analysis, qualitative adjustment, reconciliation and dissemination, name the owner and the output of each.
• ▸Forecasting champion: name the person and note their experience and training for managing the function.
• ▸Accuracy metrics: list the measures used and the acceptable-accuracy threshold per product group.
• ▸Cost record: enter management, operations and marketing costs, and the desired customer service level behind the threshold.
• ▸Technique and parameter log: record the technique per product, the parameter values used, and the tested rationale for each.
• ▸Exception list: list products the system did not bring to acceptable accuracy, and who reviews them.
• ▸Adjustment documentation: for each qualitative adjustment, record who made it, the reasoning, and whether it improved accuracy.
• ▸Training status by role: record what maintainers, adjusters and users have been trained in, and what is outstanding.
• ▸Derived-demand check: note any demand spikes caused by a partner's ordering policy and the demand-planning response considered.