When Water-Cut Data Goes Stale, Prediction Is Not Measurement

Water cut is the fraction of produced liquid that is water rather than oil. Once the component volumes are known, the calculation is simple:

water cut = water volume / (oil volume + water volume).

Knowing the current value is harder.

An offshore completion may have a valid reported value for one period and a materially different value later. Between observations, an operator can retain the last value, extrapolate a trend or use a model. Each can be useful. None converts an estimate into a contemporaneous measurement.

That distinction matters because water production affects separation, handling, treatment and disposal requirements. Public US Department of Energy material also identifies corrosion, scale, separation cost and eventual loss of economic production among the consequences associated with high produced-water fractions. US DOE National Energy Technology Laboratory

The practical question is therefore not whether forecasting has value. It is:

How quickly can a previously observed water cut cease to be an adequate description of the current production state?

We examined that question using only publicly traceable US federal offshore production records.

The evidence source

The Bureau of Safety and Environmental Enforcement publishes annual delimited files from the Oil and Gas Operations Report, Part A, or OGOR-A. The files provide monthly well-completion production records from 1996 onward. BSEE OGOR-A data

The associated field definitions identify, among other items:

• the completion name;
• production month;
• days on production;
• monthly oil volume;
• monthly water volume; and
• the API well number.

The oil and water fields are monthly quantities reported for a completion, not instantaneous flow measurements. BSEE OGOR-A field definitions

The reporting basis is important. ONRR's OGOR-A instructions require a separate report each month for leases or agreements containing active wells, unless non-monthly reporting has been approved. The information is collected to corroborate production and disposition data with sales and royalty data. ONRR Form 4054-A and instructions

This makes OGOR-A useful for a longitudinal census of reported production. It does not make the database an instrument-level reference for real-time water cut.

What we analysed

We analysed the BSEE OGOR-A annual files for 2010 through 2019, downloaded on 12 June 2026.

A completion string was identified by the combination of API well number and completion name. We retained rows with a positive number of days on production and constructed monthly water cut from the reported oil and water volumes.

To reduce instability from very small liquid denominators, a month qualified only when:

• reported oil plus water was at least 1,000 barrels for the month; and
• reported oil volume was greater than zero.

A completion entered the longitudinal cohort only if it had at least 12 qualifying months. Those months did not have to be consecutive.

The downloaded files contained 7,128 completion identifiers with production records. After applying the volume, oil-positive and minimum-history criteria, 4,686 completion strings remained.

For each retained completion, we compared months separated by exactly one, two, three and six calendar months. The reported statistic is the absolute difference in water cut between the two months, expressed in percentage points.

This creates a direct stale-value test. If the earlier month's water cut were held unchanged until the later month, the absolute change would also be the absolute error of that held value against the later monthly report.

Five- and ten-percentage-point differences were selected as descriptive thresholds. They are not presented as universal operating tolerances; the acceptable error for a real decision depends on the asset, measurement system, economics and intended use.

What the records show

Across the 4,686 retained completion strings, the median observed water-cut range was 46 percentage points. Here, “range” means the maximum minus the minimum monthly water cut among the qualifying observations available for that completion between 2010 and 2019. It is not a claim about monotonic change, full productive life or the path taken between those values.

The distribution was broad:

• 62% of retained completion strings had an observed range of at least 30 percentage points.
• 47% had an observed range of at least 50 percentage points.

The exact-gap comparison showed:

• One month: 302,782 pairs; median absolute change 1.0 percentage point; 15% changed by at least 5 points and 6% by at least 10.
• Two months: 293,512 pairs; median absolute change 1.4 points; 21% changed by at least 5 points and 10% by at least 10.
• Three months: 286,588 pairs; median absolute change 1.7 points; 24% changed by at least 5 points and 12% by at least 10.
• Six months: 270,381 pairs; median absolute change 2.3 points; 31% changed by at least 5 points and 16% by at least 10.

Counts are exact for the stated cohort. Medians are rounded to one decimal place and proportions to the nearest whole percentage.

The central result is not that every completion changes quickly. The medians are modest. The result is that the distribution has a consequential tail: even at a one-month interval, tens of thousands of matched pairs differed by at least five percentage points. At six months, nearly one pair in three crossed that descriptive threshold.

A median can conceal the operational problem

A single summary number cannot describe this distribution adequately.

If the median one-month change is 1.0 percentage point, it may be tempting to conclude that holding the last value is usually safe. That conclusion does not follow. The median describes the middle pair. It does not identify which completion-month pairs belong to the larger-change tail before the later observation arrives.

This is the operational asymmetry. After both monthly reports are available, identifying a stale value is trivial. Before the later report is available, the system must decide whether the inherited value remains adequate, whether an estimate is supportable or whether fresh evidence is required.

The age of the last observation is therefore necessary information, but age alone is not a sufficient uncertainty model. Two values of the same age may have very different reliability because the underlying completions may have different histories, interventions, flow regimes, measurement conditions and data quality.

Last observation held is a baseline, not a description of industry practice

This analysis does not claim that offshore operators universally hold the previous month's water cut unchanged. Production surveillance and allocation practices vary, and operators can use well tests, meters, engineering models and reconciliation processes unavailable in the public OGOR-A files.

The held-value calculation is used because it answers a clean, reproducible question:

If no information beyond the previous reported water cut were used, how far would that value be from the next report?

It is a baseline for quantifying observation staleness—not a complete representation of field operations.

More sophisticated prediction can improve on that baseline. Whether it does so must be demonstrated through time-ordered evaluation against later measurements, with no use of future information and with the intended operating boundary declared in advance.

The BSEE census does not perform that model comparison. It therefore cannot establish that surface history, machine learning or any other model class is incapable of estimating current water cut. It shows only that the previously reported value often ceases to be an adequate substitute for the later monthly value under the stated thresholds.

Prediction and measurement answer different questions

A measurement asks what the available physical observation supports at a particular time and under a defined measurement process.

A prediction asks what value is plausible given prior observations, present inputs, model assumptions and calibration.

The two can agree closely. They can also diverge while each remains internally consistent: the measurement may be noisy or aggregated, the prediction may be outside its validated operating regime, or the physical system may have changed in a way not represented by the model's inputs.

Presenting a prediction as if it were measured removes information that the decision-maker needs. A professional estimate should retain:

• the time to which it applies;
• the age and source of its latest physical anchor;
• the data channels available at that time;
• the operating conditions represented in validation;
• its uncertainty or decision interval;
• any reason the estimate should be rejected; and
• the distinction between estimated, measured and reconciled values.

This is not a semantic preference. It determines whether downstream users can audit the number and decide what level of action it can support.

What a credible evaluation must report

A water-cut estimation result should be judged as an operating system, not as a single average error.

At minimum, a public evaluation should state:

• Reference: what produced the comparison value and whether it was a monthly allocation, well test, meter output or another reference.
• Unit of analysis: well, completion string, test, time interval or field.
• Cohort: every inclusion and exclusion criterion, including minimum flow or liquid-volume thresholds.
• Time boundary: how development, calibration and evaluation periods were separated.
• Leakage control: whether any information from the evaluation future entered training, calibration, feature construction or data cleaning.
• Baselines: at least the last measured value and other incumbent estimates relevant to the intended operation.
• Coverage: how often the system issued an estimate and how often it declined because evidence was insufficient.
• Operating conditions: the regimes, equipment configurations and missing-data states represented in testing.
• Transfer: performance on completions, fields, periods or facilities not used to construct the system.
• Error distribution: not only a mean or median, but tail errors at decision-relevant thresholds.

If a system is recalibrated for every evaluation asset, the result demonstrates local fitting rather than unassisted transfer. If rejected cases disappear from the denominator, reported accuracy does not describe deployed coverage. If monthly averages are treated as instantaneous truth, the reference claim is stronger than the data support.

These distinctions must remain visible.

Limits of the BSEE census

This study has several explicit limitations.

First, OGOR-A contains monthly reported production volumes. It cannot determine when within a month a change occurred, whether the change was gradual or abrupt, or how much shorter-timescale variation was averaged into the monthly total.

Second, the pair-level percentages are descriptive. A completion can contribute many pairs, so the pairs are not statistically independent observations of separate wells. We do not present conventional confidence intervals that would incorrectly assume otherwise.

Third, the liquid-volume threshold and requirement for positive oil production exclude low-volume and zero-oil months. The results should not be extended to those states without a separate analysis.

Fourth, the 2010–2019 window truncates each completion's history. The reported max–min value is an observed range inside that window, not a lifetime water-cut change.

Fifth, five and ten percentage points are analysis thresholds, not universal decision limits.

Sixth, the database is assembled from documents submitted to the Federal Government. BSEE warns that errors may exist in its working database and that the computer data is not a legal document. BSEE Data Center disclaimer

Finally, this is one public US offshore dataset. It does not establish performance for another basin, operator, allocation system, measurement architecture or production regime.

What the evidence proves—and what it does not

The analysis supports three bounded conclusions.

First, water cut is not a fixed completion property in this cohort. Large observed ranges were common over the retained histories.

Second, holding the previous monthly water cut produced a non-trivial tail of errors against later monthly reports, and that tail became larger at longer exact gaps.

Third, an estimate should carry its evidence age and status. A number inherited from an earlier observation is not a current measurement merely because it remains in a database or operating model.

The analysis does not validate a Mondren estimation system. It does not compare forecasting algorithms. It does not establish that all history-based or surface-signal approaches must fail. It does not prove real-time accuracy, transfer to a new field, production improvement or economic benefit.

Those claims require separate evidence.

The Mondren perspective

At Mondren, water-cut estimation is a frontier research area. Our public position is deliberately narrow:

A current water-cut claim should be treated as an estimate with a known evidence age and declared validity—not as a permanently valid property inherited from the last observation.

A credible solution must be anchored to independent physical evidence, tested without future leakage, evaluated across transfers and allowed to refuse an estimate when the available information does not distinguish the relevant operating states.

This principle can be stated publicly without disclosing our model construction. Our internal representation, estimator, signal transformations, calibration procedure, correction logic, parameter values and deployment gates remain proprietary.

The commercial implication

The commercial opportunity is not to replace every measurement with a prediction. It is to reduce the period during which important operating decisions depend on an unqualified stale value.

That can support better surveillance, prioritisation and planning between direct observations. The value must ultimately be demonstrated in the operator's own workflow: against its accepted reference measurements, at its required update frequency, within its measurement architecture and under its economic error tolerances.

A deployable product should make three states unmistakable:

1. Measured: supported by a declared physical measurement process.
2. Estimated: inferred from current and historical evidence within a validated boundary.
3. Unavailable: not supportable with sufficient confidence under present conditions.

The third state is as important as the first two. A system that always emits a number can create apparent continuity while concealing loss of evidence.

Water-cut data becomes operationally dangerous not simply when it is old, but when its age, uncertainty and provenance disappear from the decision. Prediction can help bridge an observation gap. It should never erase the fact that the gap exists.

This article reports an analysis of public BSEE OGOR-A production records. It does not make a claim of field validation, deployment readiness, production improvement or economic benefit for any Mondren system.