From Verification to Forecasting: Monitoring Official Crime Trends in Bangladesh

Crime is one of the biggest challenges facing Bangladesh, where more than 170,000 incidents such as theft, robbery, kidnapping, murder, and violence against women and children are officially recorded every year. Bangladesh Police Headquarters (PHQ) publishes these counts every month, but this national data is only stored and reported, never used to look ahead and prepare. This paper builds a system that does exactly that. We collected 88 months of officially reported crime data (January 2019 to April 2026, 17 police units, 15 crime categories) and linked every number back to its public source. We then trained seven common forecasting models and a combined ensemble, and checked them two ways. First, we forecast crime 1, 3, 6, and 12 months ahead using only older data and measured the error. Second, we waited for the police to publish the real numbers for those months and compared them with our earlier forecast. For the national monthly total, our forecast matched the published numbers with 98.9, 97.8, 92.8, and 86.1 percent accuracy at 1, 3, 6, and 12 months, and a rolling test over many starting points stayed between 92.6 and 89.1 percent on the steady categories. Because this check repeats every month as new data arrives, we built it into a working application that scores the previous forecast, adds the new month, retrains, and watches for drift, giving continuous monitoring instead of a one-time study. We forecast only reported crime and make no claim about crimes that are never reported.