The Modified Arps decline curve

The Arps equations

J. J. Arps formalized the decline curve family in 1945, and it remains the backbone of empirical production forecasting. The general hyperbolic form gives the production rate q at time t from the initial rate qi, the initial nominal decline Di and the exponent b:

q(t) = qi / (1 + b · Di · t)^(1/b)

Two special cases fall out of that single equation. When b = 0 it reduces to exponential decline, q(t) = qi · exp(−Di · t), in which the decline rate never changes. When b = 1 it becomes harmonic decline, q(t) = qi / (1 + Di · t), the slowest-flattening member of the family.

Everything between those bounds is hyperbolic decline, which is what most producing wells are fitted with.

What each parameter controls

The three parameters are not interchangeable knobs — each moves a different part of the curve, which is why all three have to be visible for a fit to be reviewable.

Di — initial decline rate

How steeply production is falling at the start of the forecast. It is fixed mainly by the most recent observed months, so it is the parameter most sensitive to what happened just before the cutoff — including downtime that has nothing to do with the reservoir.

b — the decline exponent

How quickly the decline rate itself decays. Larger b means the curve flattens faster, so the well produces longer at low rates and the extrapolated volume grows. Because b acts on the tail, small changes in b produce large changes in EUR while barely moving the near-term forecast.

Terminal decline

The floor. Once the hyperbolic decline rate falls to this value, the forecast continues exponentially at that constant rate. It has almost no effect on the first years of the forecast and decisive effect on the total volume.

Why the terminal decline exists

This is the failure the modification is designed to prevent, and it is worth stating precisely because it is invisible in the first few years of a forecast.

In hyperbolic decline the decline rate decreases with time. For b ≥ 1 the resulting curve flattens slowly enough that its integral over infinite time does not converge — the cumulative volume grows without limit. Even for b comfortably below 1, a long-lived well can accumulate an implausible tail, because the model has no notion of the well eventually becoming uneconomic.

Imposing a terminal decline caps that behaviour: below a defined minimum decline rate the curve becomes exponential, which always integrates to a finite volume. In a batch forecast across hundreds of wells this is not a refinement — it is the difference between a portfolio EUR that means something and one inflated by tails nobody inspected.

Fitting the model to real production

The fit finds the parameter set that best reproduces the observed history, but the observed history is not a clean signal. Rate changes caused by choke adjustments, well service, facility downtime or allocated volumes after a restart all look like decline behaviour to a curve-fitting routine.

PronoHorizon keeps those events marked rather than smoothing them away: months with no production report stay missing, allocated volumes are flagged as estimated rather than reported, and days-on counts stay attached so a partial month is not read as a rate collapse. The fit is then something an engineer can argue with, because the artefacts are visible.

The same three parameters produce the P10, P50 and P90 cases. Editing them moves all three cases together, since the range describes uncertainty about one decline, not three independent forecasts.

What an empirical decline model cannot do

Arps decline is a curve fitted to rate history. It carries no reservoir physics, no material balance, no drive-mechanism reasoning. It cannot anticipate a workover, a compressor installation, an infill well, a change in operating strategy, or a shut-in, because none of those events are present in the history it was fitted to.

That is not an argument against using it — it is the reason it needs both engineering review and holdout validation around it. The parameters give the engineer something to correct; the backtest measures how often correction was needed.