Forecasting OCTG Demand from Rig Schedules

Method

We pulled state permit feeds (Texas RRC, New Mexico OCD, Oklahoma OCC) and overlaid them with publicly available rig schedules and our consignment consumption history. The hypothesis: stockout events are not random — they follow rig activity with a predictable lag.

The model is intentionally simple. A linear regression on three features:

1. Permit count (rolling 14-day, by basin)

2. Active rig count (current week, by basin)

3. Historic consumption (rolling 8-week, by SKU class)

Output: probability of stockout per SKU class per yard, two weeks ahead.

Result

Across the SKUs we tested, the model flagged 78% of actual stockout events 14 days before they happened. False positive rate sat around 12%, which we tuned against by raising the threshold for high-cost classes.

The miss rate was concentrated in two places: surprise rig moves (an operator skidding to a new pad without updating their public schedule) and emergency well-control consumables (which are by definition unpredictable). For those we keep a separate buffer.

What changed in the field

Once the model was running, our procurement team started getting a Monday morning report: a ranked list of SKU/yard combinations likely to stock out within 14 days. They could pre-position parts before the demand hit.

Hot-shot delivery costs across the consignment program dropped roughly 60% in the quarter following deployment. That paid for the data engineering inside the first six weeks.

Why we're not selling this as software

A few reasons. The model only works because it sits on top of TIBRSupply's actual consumption history — without the inventory data, the permit and rig features alone are noisy. We'd rather embed it in the consignment program than try to package it.

That said: if you're running your own inventory and want to talk about adapting the approach, we will. The math is not a secret. The data plumbing is the hard part.

Caveats

• Permit data lags reality by a few days in some states. We compensate by inflating the rolling window.
• The model is basin-specific. A model trained on the Permian does not transfer cleanly to the Bakken — different rig activity patterns, different SKU mix.
• It is not a replacement for human judgment on long-lead engineered items. Those still get planned manually against the well program.