Time-series algebra is littered with lagged terms: x_{t-1}, x_{t-2}, differences like x_t - x_{t-1}, sums of them stretching back years. Writing all those subscripts by hand is not just tedious — it hides the structure. The backshift operator (also called the lag operator) is a single piece of notation that turns those tangled subscripts into ordinary algebra. Once you have it, an autoregression, a difference, and a seasonal filter all become polynomials you can multiply, factor and cancel like numbers.

The whole idea is one definition. The operator B takes a series and shifts it one step into the past:

That is all B does — reach back one tick. Apply it twice and you reach back two, and in general

By convention B^{0} = 1 is the "do nothing" operator (1\cdot x_t = x_t). Some authors write L for "lag" instead of B for "backshift" — same operator, different letter.

Treat B like a number

The quiet miracle is that B obeys all the ordinary rules of algebra. It is a linear operator, so it distributes over sums and pulls out constants:

Powers of B add the way exponents should — B^{i}B^{j} = B^{\,i+j} — and different powers commute, because shifting back 2 then 3 is the same as shifting back 3 then 2. So any expression built from B is just a polynomial in B, and you may add, multiply and factor those polynomials exactly as if B were a variable like z. This is the licence that makes the whole ARMA machinery run.

Seeing the shift

Applying B geometrically means sliding the whole series one tick to the right: the value that used to sit at time t-1 now lines up under time t. Below, the solid line is x_t and the dashed line is its backshifted copy B x_t. They have the same shape — one is just the other, delayed by one step.

Reading a lagged model is now reading a delay: every extra power of B nudges a copy one more step to the right.

Differencing, in one symbol

The operation you already know from differencing — subtracting each value from the next — is just 1 - B. We give it its own symbol, the difference operator \nabla:

Differencing twice is \nabla^2 = (1-B)^2 = 1 - 2B + B^2, which expands to x_t - 2x_{t-1} + x_{t-2} — the polynomial algebra doing the bookkeeping for you. A seasonal difference of period s (e.g. s=12 for monthly data) reaches back a whole season:

Trend removal, seasonal removal, and their combinations are now just products of these factors — a fact the seasonal ARIMA models lean on completely.

Writing whole models compactly

Here is the payoff. An autoregression on p lags,

collects all its lag terms into one polynomial. Move them to the left and factor out x_t:

The bracket is the AR polynomial \phi(B). A moving-average model does the same on the noise side with an MA polynomial \theta(B) = 1 + \theta_1 B + \dots + \theta_q B^{q}, and the general ARMA model is the elegant one-liner

Whole families of models now live in two short polynomials. And because they are polynomials, you can factor them: the roots of \phi(z)=0 and \theta(z)=0 decide whether a model is well-behaved — the subject of causality and invertibility.

Worked example

Expand (1 - B)(1 + B)\,x_t. Multiply the polynomials just as you would numbers: (1-B)(1+B) = 1 - B^2 (the cross terms +B and -B cancel). Apply it:

So differencing "up then down" collapses to a single lag-2 difference — a fact you would have sweated over in raw subscripts, delivered instantly by the algebra.

A second one: (1 - 0.5\,B)\,x_t. This is the AR(1) polynomial with \phi = 0.5. Distribute:

Setting that equal to \varepsilon_t recovers x_t = 0.5\,x_{t-1} + \varepsilon_t — the familiar autoregressive model, now packaged as a single factor 1 - 0.5B.