What it means
The method has three named stages that run as a loop. Identification chooses how much of the past to use, estimation fits the numbers, and diagnostic checking asks whether the unexplained remainder still contains a pattern.
If it does, you go back and change the model rather than accepting the forecast. An ARIMA model has three settings, usually written as p, d and q.
The p is how many past values feed into the forecast, the d is how many times the series must be differenced, meaning converted into period-to-period changes to strip out a trend, and the q is how many past forecast errors are carried forward. A plain ARIMA with p of 1 and nothing else is simply a forecast built from last period's value.
Differencing is the step that non-statisticians find odd, but it does the heavy lifting. Most business series grow over time, and the method needs a series whose average and variability stay roughly constant, so you model the changes rather than the levels.
Seasonal versions apply the same trick across years, for example by comparing each month with the same month a year earlier. In practice the method works best on a reasonably long, stable history with no outside drivers you need to explain.
Forecasting next quarter's electricity demand from 10 years of monthly data suits it well, while forecasting the effect of a price change does not, because the cause is not in the series' own history. For that, a model containing explanatory variables is the right tool.
The honest limitation is that it assumes the future behaves like the past. A Box-Jenkins forecast will not see a new competitor, a regulatory change or a one-off disruption coming, and it will happily project a trend straight through a turning point.
Most finance teams use it as a baseline and then adjust for what they know that the data cannot.
In practice
Real-world examples.
Example
A utility forecasts monthly electricity demand from 12 years of history using a seasonal version of the method. The forecast drives fuel purchasing, and the team compares each month's actual figure against the model's prediction range to spot genuine changes in consumption.
Example
A hospital group models weekly emergency department attendances to plan staff rotas. Because attendance has a strong weekly and annual pattern but no obvious commercial driver, the method fits well, and the model is re-estimated every quarter with the newest data.
Example
A treasury team builds a Box-Jenkins forecast of daily cash receipts to decide how much to leave in its overnight account. The baseline is then overridden manually for known events such as one large customer settlement, which is how most teams combine a statistical model with human knowledge.
Formula
Calculation
For the simplest autoregressive version, the forecast is: next value = constant + (coefficient x most recent value)
The long-run average the series settles towards is: constant / (1 - coefficient)
Suppose monthly service revenue, measured in thousands of dollars, is fitted with a constant of 200 and a coefficient of 0.6. The most recent month was 1,000, so the forecast for next month is 200 + (0.6 x 1,000) = 800. The month after that uses the new figure, 200 + (0.6 x 800) = 680, and the one after that gives 200 + (0.6 x 680) = 608. The series is converging on a long-run average of 200 / (1 - 0.6) = 200 / 0.4 = 500, which tells the finance team that the recent month of $1,000,000 was unusually strong rather than a new normal.Case study
Seen in the real world.
Varmont Logistics is an illustrative, fictional freight company that budgeted next year's volumes by taking last year's figures and adding 5% across the board. Monthly variances of 20% or more were routine, and the operations team had stopped trusting the budget.
A new financial analyst fitted a seasonal Box-Jenkins model to seven years of monthly tonnage. The diagnostic stage showed a clear annual pattern that the simple growth rule had ignored, with a peak each autumn and a trough in midsummer, and the fitted model cut average monthly forecast error from about 18% to about 7%.
Varmont kept the statistical forecast as its baseline and added a manual overlay for known contract wins and losses, which no model could see in the history. The illustrative lesson is that the method improved the shape of the forecast while judgement was still needed for anything genuinely new.
Watch out
Common mistakes.
- Fitting a model to a short or unstable history, when the method needs a reasonably long run of comparable periods to say anything useful.
- Skipping the diagnostic stage and accepting the first model that fits, which is how a pattern left in the residuals quietly becomes a forecasting error.
- Using the method to answer cause-and-effect questions such as the impact of a price rise, which needs a model containing that explanatory variable.
Questions
People also ask.
What do the letters p, d and q actually mean?
They are counts: p past values used in the forecast, d rounds of differencing to strip out a trend, and q past forecast errors carried forward.
Is this better than a simple moving average?
It usually is for a series with trend and seasonality, because it fits those features explicitly, though for a short or very noisy history a simple average can be more reliable.
Do I need specialist software?
Any standard statistical package and most spreadsheet add-ins will fit these models, and the harder part is the judgement at the identification and diagnostic stages rather than the arithmetic.
From the founder's library
Take it further with the book.
Build your financial confidence beyond this definition. Shihan's full-length guide, Accounting Fundamentals, takes the same plain-English approach and turns it into a complete, practical playbook for non-finance managers, business owners and students - with chapter-end quiz answers and presentation slides included.
25% off with code MMHQ25, applied at checkout. Priced in USD - checkout may show the equivalent in your local currency.
View the book and save 25%Related