Predictive Analytics

Cold-coffee concentrate has to be ordered on Thursday for the following week. Order too little and Nisha turns customers away; order too much and it spoils. She does not need to know exactly what will happen. She needs a good estimate, and an honest sense of how wrong it might be.

The rule

The past, patterned, projects the future, with a margin.

Predictive analytics finds a pattern in history and extends it forward. The patterns analysts lean on most:

• Level: the usual amount. Wednesdays at Kettle & Crumb sell around 37 cups.
• Trend: a steady drift up or down, like weekly totals rising 10 cups a week through spring.
• Seasonality: a repeating cycle, such as weekends beating weekdays or summer beating winter.
• Events: one-offs like the fair, which must be handled separately so they do not distort the normal pattern.

Methods range from a simple average of recent weeks to regression and machine-learning models. Whatever the method, every forecast rests on one assumption: the future will behave like the past, unless you know something has changed.

Here are the last four Wednesdays:

The shortcut

A forecast has three parts, and a forecast missing any of them is incomplete: a number, a range, and the assumption behind it.

The traps: forecasting from a single week, letting one event day inflate every prediction, and giving one confident number with no range. A forecast of "exactly 37" will almost certainly be wrong; "34 to 40" is useful.

Drills

Answer fast — trust the rule.

Using the four Wednesdays in the table, what is the average level?

Remember

• The past, patterned, projects the future, with a margin.
• Shortcut: every forecast needs a number, a range and its assumption.
• Traps: one week of history, event days left in, and no range.