# Forecasting Future Murders of Mr. Boddy by Numerical Weather Prediction

## TL;DR

This paper applies variational data assimilation to model and forecast the timing and details of Mr. Boddy's murders, incorporating new parameters like motive and health condition to improve prediction accuracy.

## Contribution

It introduces a novel approach using data assimilation with expanded parameters to predict murder events in a complex, multi-factor scenario.

## Key findings

- Forecasted next murder time with a 7-hour standard deviation
- Identified key parameters influencing murder prediction, including motive and health
- Discussed limitations related to Gaussian error assumptions and landscape complexity

## Abstract

Despite a previous description of his state as a stable fixed point, just past midnight this morning Mr. Boddy was murdered again. In fact, over 70 years Mr. Boddy has been reported murdered $10^6$ times, while there exist no documented attempts at intervention. Using variational data assimilation, we train a model of Mr. Boddy's dynamics on the time series of observed murders, to forecast future murders. The parameters to be estimated include instrument, location, and murderer. We find that a successful estimation requires three additional elements. First, to minimize the effects of selection bias, generous ranges are placed on parameter searches, permitting values such as the Cliff, the Poisoned Apple, and the Wife. Second, motive, which was not considered relevant to previous murders, is added as a parameter. Third, Mr. Boddy's little-known asthmatic condition is considered as an alternative cause of death. Following this morning's event, the next local murder is forecast for 17:19:03 EDT this afternoon, with a standard deviation of seven hours, at The Kitchen at 4330 Katonah Avenue, Bronx, NY, 10470, with either the Lead Pipe or the Lead Bust of Washington Irving. The motive is: Case of Mistaken Identity, and there was no convergence upon a murderer. Testing of the procedure's predictive power will involve catching the D train to 205th Street and a few transfers over to Katonah Avenue, and sitting around waiting with our eyes peeled. We discuss the problem of identifying a global solution - that is, the best reason for murder on a landscape riddled with pretty-decent reasons. We also discuss the procedure's assumption of Gaussian-distributed errors, which will under-predict rare events. This under-representation of highly improbable events may be offset by the fact that the training data, after all, consists of multiple murders of a single person.