Abstract

Predicting the future production of Proved UnDeveloped (PUD) reserves is an important part of the evaluation of any asset. In the absence of production history, the most reliable way to make a PUD forecast is by analogy with production histories from similar nearby completions. This is most often done using a type curve. A traditional type curve is generated by normalizing to one the production histories of analogous wells and finding the shape of a curve which best fits their average decline. This is a tedious process when done by hand, and results are so difficult to test that in practice they rarely are. Our study introduces a methodology for testing statistical type curves based on 5-fold cross validation (5FXV), which rigorously evaluates the ability of an algorithm to produce accurate type curves, giving confidence that those type curves can be used to make careful and informed decisions. We offer two case studies to demonstrate the strength of the method and introduce a useful extension of the method.

A physio-statistical algorithm named BetaZi (BZ) automatically generates type curves which are considerably richer than a simple shape. They include information about actual production ranges abstracted as percentiles (90% of the time, actual production from a new well is expected to exceed the BZ p90 bound; 50% of the time it should exceed the p50, and so on.) 5-fold cross validation (5FXV) is used to test the performance of the algorithm. In 5FXV, a random 10% of the data from analogous wells is held out for testing and the remaining 90% is used to generate a type curve. A quantile-quantile (Q-Q) evaluation is made which compares the number of times the test data exceeded the bounds predicted by the type curve. The process is repeated 5 times on different subsets of data.

For this paper, we use 5FXV to estimate type curves for several plays in the Bakken and Kansas. The results show that excellent performance can be achieved if enough analogous wells are available to generate the type curve. Type curves that are made with wells that are not analogous can lead to overinflated estimations of value, as is shown with the Kansas wells. Finally, a p-score is computed for each well using the type curve, allowing production from wells of different histories to be conveniently plotted on one map or cross-plotted with other parameters.

You can access this article if you purchase or spend a download.