AI Glossary / Time series and signal processing / Intermediate

Backward shift operator

Definition

The backward shift operator is a notation that moves a time-series value back by one observation period. Written B, it satisfies B y_t = y_(t-1); applying it k times gives B^k y_t = y_(t-k). It makes lagged relationships, differencing and time-series model equations more compact and easier to manipulate.

What it means

The operator changes the time index, not the numerical size of the observation. If y_t is this month's value, B y_t refers to last month's value. B² refers to two periods earlier, not to squaring the previous observation. Some texts use L for the same lag operator.

This notation is especially helpful when a model contains many lags or combines ordinary and seasonal differences. It expresses a reusable operation on the whole sequence. The meaning of one period depends on the data frequency, so a lag of twelve represents a year for monthly observations but not for every dataset.

How it works

Treat the symbol 1 as the identity operator, which leaves the series unchanged. Then (1 - B)y_t means y_t - y_(t-1), the first difference. Applying that difference twice gives (1 - B)²y_t = y_t - 2y_(t-1) + y_(t-2).

This differs from (1 - B²)y_t, which subtracts the observation two periods earlier. With constant coefficients, operator polynomials can be expanded using ordinary algebra. For example, (1 - B)(1 - B^m) combines a first difference with a lag-m difference. Observations at the required earlier indices must be available, and boundary values need explicit handling. These algebraic identities do not by themselves justify which transformation is statistically appropriate.

Three practical examples

1. Forecast specification: An analyst writes an autoregressive model compactly as (1 - phi B)y_t = e_t. Expanding it clarifies that the previous observation enters with coefficient phi.

2. Seasonal preprocessing: A monthly-data team uses (1 - B^12) to express the difference from the same month one year earlier. It retains the calendar grid so the lag has the intended meaning.

3. Implementation review: A developer checks that software applies a second ordinary difference rather than a single two-period difference. The operator expressions expose a mistake that similar variable names could conceal.

Think of it this way

Think of B as moving a bookmark one place backward in an ordered record. Moving it twice selects an older record; it does not alter that record's numerical value. Expressions such as 1 - B then specify how the current and selected earlier records should be combined.

Worked example

Consider an original series with four consecutive values: 10, 13, 17 and 22. At the fourth observation, y4 = 22, B y4 = y3 = 17 and B²y4 = y2 = 13.

The first difference is (1 - B)y4 = 22 - 17 = 5. The second ordinary difference is (1 - B)²y4 = 22 - 2 × 17 + 13 = 1. This agrees with differencing the two latest first differences: (22 - 17) - (17 - 13) = 5 - 4 = 1.

A single two-period difference instead gives (1 - B²)y4 = 22 - 13 = 9. The values 1 and 9 answer different questions. For another check, (1 - 0.5B)y4 = 22 - 0.5 × 17 = 13.5. These calculations illustrate operator application only; they do not fit an autoregressive model or determine whether differencing is appropriate for this very short sequence.

Illustrative case study

Fictional case: Riverbend Forecasting compares a documented demand model with a newly implemented pipeline. The documentation specifies second-order differencing, but the pipeline subtracts the value two weeks earlier. Forecasts disagree even though both teams describe their transformation as a second difference.

The reviewer expands the two operator expressions and identifies the missing coefficient of minus two on the one-period lag. The team corrects the transformation, adds a small independently calculated reference case and documents the lost initial observations. It then reruns model fitting and forecast evaluation, because changing preprocessing can alter both parameter estimates and how forecasts are returned to their original scale.

Common misconceptions and mistakes

B²y_t is not (B y_t)². The exponent counts shifts. Likewise, (1 - B)² is not 1 - B²: the cross-term -2B matters. A lag-m difference and m successive ordinary differences are generally different transformations.

The operator does not fill missing data or guarantee equal time spacing. If records are removed, shifting one row may no longer mean shifting one calendar period. Backshift notation also does not make a nonstationary series stationary automatically. Formal inverse-operator expressions require appropriate conditions and initialization; they should not be treated as unrestricted numerical division by a symbol.

Business use

The main business benefit is clearer communication between statistical models and their software implementations. Compact equations can make forecasting specifications easier to audit, particularly when multiple lag structures or seasonal components are involved.

Require implementations to state the observation frequency, lag convention, initial-value handling and transformation order. Check sample calculations at boundaries and around missing dates. When forecasts are produced on a differenced scale, document how they are accumulated back to levels. These details can materially change a planning output even when the model's name remains unchanged.

People also ask

Does B always mean one day?

No. It means one period at the series' stated observation frequency.

Can B be applied more than once?

Yes. B^k shifts back k periods.

Is (1 - B²) the second ordinary difference?

No. It is a two-period difference; the second ordinary difference is (1 - B)².

Do ordinary and seasonal differencing commute?

With the usual constant-coefficient operators, their algebraic products commute where the required observations are defined. Boundary and missing-data implementation choices still require care.

Check your understanding

What value does B² select in the example at period four?

Answer: It selects 13.

What is the second ordinary difference there?

Answer: It is 1.

What does (1 - B²) produce instead?

Answer: It produces 9, the difference between values two periods apart.

Related terms

• Lag operator
• Differencing
• Seasonal difference
• ARIMA model
• autoregressive polynomial
• difference equation
• time-series transformation

Learn next

Learn next how backshift polynomials express ARMA and seasonal ARIMA models. Practice expanding the operators and reversing differences with known initial values before relying on compact notation in a software specification.

More in Time series and signal processing

• Autocorrelation - Autocorrelation is the correlation between a variable and a time-shifted version of itself.
• Autocorrelation function - The autocorrelation function, or ACF, gives a time series' autocorrelation across different lags.
• B-splines - B-splines are basis functions used to represent piecewise-polynomial curves and surfaces.
• Box-Jenkins methodology - The Box-Jenkins methodology is an iterative approach to building time-series models, especially ARMA and ARIMA models, through identification, parameter...
• Box-Pierce test - The Box-Pierce test is a portmanteau test that assesses whether a sequence has autocorrelation across a chosen set of initial lags.
• Classical decomposition method - The classical decomposition method separates a time series into estimated trend-cycle, seasonal and remainder components using moving averages and seasonal...

Sources

Hyndman and Athanasopoulos, Forecasting: Principles and Practice, Backshift notation: https://otexts.com/fpp2/backshift.html . The author-maintained reference explicitly distinguishes repeated ordinary differencing from lagged differences. The original numerical examples were independently checked. The older supplied glossary's potentially confusing second-difference wording is resolved using the explicit operator expansions.

Previous: Backward elimination Next: Bacterial colony optimization