From Arps to AI: Lessons from the SPEE Software Symposium

The Society of Petroleum Evaluation Engineers (SPEE) recently held its annual Software Symposium, this year titled "Experiences in Automated Forecasting". It brought together industry professionals to discuss trends, challenges, and advancements in automated forecasting (AF). Among the many presentations, three stood out: survey results on AF adoption, a historical perspective on Decline Curve Analysis (DCA), and a "bake-off" of automated forecasting tools. Here’s a closer look at these key highlights.

Survey Results: How is Automated Forecasting Being Used?

Symposium chairman Lucas Smith reviewed the results of a survey on AF usage within the industry. The findings shed light on how companies are incorporating these tools into their workflows. About half of the respondents reported using AF, but there’s still a level of caution. While AF is primarily applied for reserves forecasting, type well profile creation, and deal screening, most users remain hands-on. They don’t simply accept the software's output; instead, they review each forecasted curve and adjust them manually. Only a small number of users trust the software enough to conduct spot checks rather than detailed reviews.

A Look Back: The History of Decline Curve Analysis

John Wright’s presentation took a historical turn, tracing the evolution of DCA from its early days. He began with the "loss ratio" method, first described in 1927, which laid the groundwork for modern approaches. This eventually led to J.J. Arps' influential 1945 paper, "Analysis of Decline Curves", which introduced the Arps equation. Even today, this formula—or its modified variations—remains the most widely used tool in the industry for analyzing production decline. Wright’s presentation was a reminder that while automation and new technologies emerge, some fundamentals continue to underpin our understanding of well performance.

The Main Event: The Bake-Off of Automated Forecasting Tools

The centerpiece of the symposium was a “bake-off” that pitted several automated forecasting software tools against one another. Each vendor was tasked with generating forecasts from truncated production data (the last 60 months removed), and their predictions were compared against actual production histories. David Fulford presented the bake-off results, with a focus on forecasting errors.

The bake-off used two key metrics to assess accuracy: median log error and the standard deviation of the log error. A few patterns emerged:

- Median Error and Basin Variability: Across all vendors, the median error hovered around zero, suggesting no consistent bias toward over- or under-forecasting on average. However, the accuracy varied significantly by basin, reflecting the unique challenges of different geological settings.

- Forecasting Complexity by Well Type: Dry gas wells proved to be the easiest to forecast, followed by oil wells. Multiphase wells presented the greatest difficulty, a challenge that aligns with real-world complexities.

- Impact of Production History: Longer production histories led to tighter forecast variances, highlighting the importance of having sufficient data before relying on AF tools. For wells with more data, the forecasts were generally more reliable.

- Tendency Toward Under-Forecasts: When the software did deviate, it often underestimated production, particularly in cases where wells had been choked back during their early life. This finding is particularly relevant for operators who manage such wells, as it suggests AF tools may need further calibration for these scenarios.

These observations were consistent with what we have seen in the field, where both manual and automated methods can struggle with the same nuances. It also underscored that while AF has made significant strides, it is far from replacing expert judgment.

Final Thoughts

The symposium highlighted a cautious but growing acceptance of automated forecasting within our industry. While AF tools are becoming more capable, they aren’t yet a plug-and-play solution. Human expertise remains critical to interpreting and adjusting forecasts, especially when dealing with complex well behaviors or limited data. The bake-off, in particular, was a reminder that even the most advanced software still requires careful oversight.