Like all other aspects of GOM3, enhancements to the Business Planning Add-on are continuous. In the summer, forecasts of the BOEM's MROV/ADV amounts were added, using random forests to forecast the likelihood that the government would accept the minimum required bid and how much would be required if their number was higher.
Since the last sale, a number of modifications to the prediction models and displays have been integrated to improve the accuracy and display of the results. Further analysis showed 95% of bids in the last 10 years have been within roughly 10 miles of active leases, so limiting our training and forecast areas has greatly increased the accuracy of a block receiving a bid. New variables were added to the model, including rejected bids and losing bids, and swapping out the density of Fields for the average Fields' Estimated Ultimate Recovery. The model was also enhanced by further decorrelating other variables.
It should also be mentioned that, for those who want to see an evaluation of currently leased blocks with the same methodology, the output now includes a second set of data to estimate bid likelihood and estimated high bids for those blocks that are currently leased.
One important item missing from the graphs above is the number of blocks in each cohort. The older model used fix cutoffs in bid likelihood, so while some of the predictions are quite high, they also only had a few blocks. The new model now sets the cohorts by the percentage of cumulative probability of all blocks to receive a bid. This is important to ensure that the output can reliably highlight the most attractive blocks to all companies, then the next most attractive, and so on. The old model presented data similarly, but was prone to highlight too few blocks that might be worth watching for more competition.
As shown above, accuracy has increased in the hindcasts for the last eight years, and significantly so for predictions of the sale in that year. The older model was good at predicting attractive blocks, but the newer model has increased accuracy in a shorter time frame.
The graphs also show accuracy decreasing when the price of oil dropped precipitously, as expected as so many fewer blocks received bids in those years. The newer model is doing a better job of not only predicting which blocks will be bid, but also of estimating the number of blocks receiving bids for the entire sale.