After differencing, a series has a lag-1 autocorrelation near -0.5. What does that suggest?

## Where the -0.5 comes from Differencing replaces each observation with the change from the previous one: `z_t = y_t - y_{t-1}`. It is the standard way to strip a trend or a wandering level out of a series so the lag structure underneath becomes readable. Now take the degenerate case: the series `y_t` is *already* white noise — pure uncorrelated draws around a fixed mean, with no trend to remove. Difference it anyway. The result is `z_t = e_t - e_{t-1}`, where `e_t` are the original independent noise terms. That differenced series is a moving-average process of order 1 with coefficient exactly -1. Work out its autocorrelation: - `Var(z_t) = Var(e_t) + Var(e_{t-1}) = 2 * sigma^2` — the variance has **doubled**. - `Cov(z_t, z_{t-1}) = Cov(e_t - e_{t-1}, e_{t-1} - e_{t-2}) = -Var(e_{t-1}) = -sigma^2`, since only the shared `e_{t-1}` term contributes. - So `r_1 = -sigma^2 / (2 * sigma^2) = -0.5`, and every autocorrelation at lag 2 and beyond is exactly zero. That is the fingerprint: **a lag-1 autocorrelation at or near -0.5, a clean cut-off after it, and a variance that went up rather than down.** For a general MA(1) with coefficient theta, the lag-1 autocorrelation is `theta / (1 + theta^2)`, which is bounded within `[-0.5, 0.5]` and hits -0.5 exactly at `theta = -1`. That boundary value is the *non-invertible* case, which is another way of saying the structure it represents is an artifact of the transformation rather than a feature of the data. ## Why it matters Over-differencing is not a free mistake: **Inflated variance.** The differenced series is noisier than the one you started with. Since forecast uncertainty flows from the residual variance, prediction intervals come out wider than they need to be. **Wasted model capacity.** The model now has to fit a moving-average term whose only job is to cancel the differencing you should not have applied. Parameters spent undoing your own transformation are parameters not spent describing the data, and they are estimated with their own error. **Estimation trouble at the boundary.** A moving-average coefficient sitting at -1 is on the edge of the invertible region. Fitting procedures behave badly near that edge — the likelihood surface flattens, estimates become unstable, and different starting points can land in different places. **Obscured structure.** Real short-lag structure in the original series gets tangled with the artificial negative correlation, making the correlogram harder to read rather than easier. ## How to confirm it No single indicator settles it. Look at several together: 1. **The variance comparison.** Compute the variance of the series before and after the difference. A useful difference reduces it; an unnecessary one increases it. If each additional difference raises the variance, the previous level was already right. This is the single most practical check. 2. **The undifferenced correlogram.** If the original ACF already decayed quickly to inside the significance bands, the series was level enough to work with and did not need differencing at all. 3. **The shape after differencing.** The over-differenced picture is specifically a large *negative* bar at lag 1 with the rest inside the bands. A large *positive* lag-1 bar with slow decay means the opposite — there is still a wandering level and more work is needed. 4. **Seasonal double-counting.** A frequent version of this mistake is applying both a seasonal difference at the seasonal period and a regular first difference when the seasonal one alone had already leveled the series. The same negative-lag-1 signature appears. ## The judgment call A sample lag-1 autocorrelation of, say, -0.42 on a short series is not a proof — sample autocorrelations wobble, and the significance band is wide when n is small. Treat the -0.5 figure as the theoretical target the over-differenced case converges on, and read a *clearly* negative lag-1 bar plus rising variance as the combined evidence, rather than hunting for the exact number. There is also a middle position worth knowing about. Slight over-differencing is sometimes tolerated deliberately: the resulting model can be more robust to a level that might start drifting later, at the cost of some efficiency now. That is a defensible tradeoff on a series whose future behaviour is uncertain, but it should be a stated choice rather than an accident, and the widened intervals should be acknowledged. ## Distinguishing it from real negative autocorrelation Not every negative lag-1 correlation is an artifact. Genuine mean-reverting behaviour — an inventory that gets restocked whenever it dips, a control system that overcorrects — produces real negative lag-1 autocorrelation in the *undifferenced* series. The distinguishing question is whether the negative bar appeared *because* you differenced. Check the correlogram before the transformation: if the negative lag-1 correlation was already there, it is a property of the process and the model should represent it. If it appeared only after differencing, and the variance rose at the same time, it is damage you caused.