Patterns and simple baselines

Put observations in order and inspect the decision frequency. Monthly totals can hide weekly stockouts, while daily movements can make a stable monthly business appear noisy. Mark holidays, promotions, closures, and distribution changes. An annotation proposes an explanation; it does not prove an event caused the observed movement. Level is a typical amount, trend is persistent movement, and seasonality repeats at a known calendar cycle. Two high months do not establish annual seasonality. More than one cycle helps distinguish recurrence from coincidence. With little history, acknowledge weak evidence and use cautious scenarios instead of declaring a stable pattern. Northstar quarterly units are twenty, thirty, forty, thirty in year one and twenty four, thirty four, forty four, thirty four in year two. Every matching quarter rises four units while the shape repeats. Seasonal naive repeats year two, predicting twenty four next first quarter. Adding another four gives a growth scenario of twenty eight, which assumes persistence rather than proving it. An additive description uses unit differences; a multiplicative description uses relative changes. Constant four-unit increments favor an additive story. If every quarter rises twenty percent, a relative story may be clearer. Neither should be chosen merely for an attractive fit. Compare forecasts issued before their outcomes and consider the business mechanism. Repeating the prior season can capture peaks a moving average would smooth away. Extending a trend can overshoot if growth slows. Treat these as competing explanations and test at the horizon matching replenishment or staffing. Keep decimals during comparison; operational rounding can be applied later under an explicit rule. Calendar alignment matters. Holidays move between weeks and periods contain different numbers of selling days. Product launches can change the customer mix and invalidate old seasonal shares. Ask which part of the old pattern should survive, which changed, and what future observations would make you revise the forecast. Naive uses the latest actual. A three-period moving average takes the latest three actuals. Simple exponential smoothing updates a level: alpha times the latest actual plus one minus alpha times the previous level. The next forecast equals the updated level. Specify the initial level so another analyst can reproduce every row. With actuals forty two, forty eight, fifty four, initial level forty two before week one, and alpha zero point five, updated levels are forty two, forty five, forty nine point five. Week-four forecasts are naive fifty four, moving average forty eight, and smoothing forty nine point five. Their differences reflect different weights, not an arithmetic contradiction. They are alternatives until validation gives grounds to prefer one. Score each week's forecast before updating its level with that week's actual. If a same-row formula uses the actual to generate the current forecast, it has seen the answer. Label origin and target beside each formula. At an origin after week three, week-four actual is unavailable, even if it appears elsewhere in your workbook. Simple smoothing without trend or seasonality gives the same level for all future horizons from one origin. Do not feed forecasts back as if they were new observations. A flat baseline is useful as a reference even when demand later changes. More complex methods should demonstrate a practical improvement over that reference. Higher alpha emphasizes the newest actual. A shorter moving window likewise responds faster but can magnify a temporary spike. Compare a small declared candidate set on historical origins and leave a final holdout untouched. A setting chosen to match a manager's preferred next number has not been fairly tested. Retain decimals when comparing forecasts. Expected demand of forty nine point five is meaningful even though shipments need whole units. Apply operational rounding separately. Also keep promotion adjustments and old backlog outside the baseline method, so users can see whether a changed plan came from the statistical update or a separate business decision. Now try these rules with your own small table from the lab. Draw the past on the left and the next week on the right. At each step, hide the rows that have not yet been seen. Ask how much weight the rule gives to the last week. Then make one week high and leave the rest as they were. Next make all the later weeks high. These two tests show why a rule that tracks a shift well may chase a short spike too hard. Keep your choice linked to the task, the lead time, and the cost of a miss.

1. Plot the relevant time scale

   Original flat diagram with separate labeled input and output boxes, using Northstar bottle icons.

   • Plot the relevant time scale
   • Year-three totals: seasonal naive 136; persistent four-unit growth 152.
   • Check units and timing
2. Compare like periods

   Original calculation card with aligned quantities and an arrow showing one change.

   • Compare like periods
   • Year-three totals: seasonal naive 136; persistent four-unit growth 152.
   • Check units and timing
3. Connect patterns to operations

   Original comparison chart with clearly labeled cases and a small assumption note.

   • Connect patterns to operations
   • Year-three totals: seasonal naive 136; persistent four-unit growth 152.
   • Keep assumptions visible
4. Three memory rules

   Original flat diagram with separate labeled input and output boxes, using Northstar bottle icons.

   • Three memory rules
   • Calculate naive, moving-average, and exponential-smoothing forecasts with correct timing.
   • Check units and timing
5. Put forecasts before observations

   Original calculation card with aligned quantities and an arrow showing one change.

   • Put forecasts before observations
   • Calculate naive, moving-average, and exponential-smoothing forecasts with correct timing.
   • Check units and timing
6. Choose responsiveness fairly

   Original comparison chart with clearly labeled cases and a small assumption note.

   • Choose responsiveness fairly
   • Calculate naive, moving-average, and exponential-smoothing forecasts with correct timing.
   • Keep assumptions visible