Below is a list of ongoing and past projects that I am working or have worked on.
Still in the works
[[90s GeoCities UNDER CONSTRUCTION.gif]]
(This is not a paid project, although with internal support from PH for the 2020 forecasts)
Inspired by Jay Ulfelder's coup forecasts from 2012 - 2015, I've tried at various points since then to also provide some coup forecasts as a public good, although not successful for all years (I'm lazy, what can I say, and I've learned that I turn into a data pedant when doing these).
- Years: 2017 - 2020
- Latest 2020 forecasts: https://www.predictiveheuristics.com/forecasts/
- WaPo article on the 2017 forecasts
PART and DemSpaces
- Years: 2018 - 2020
IARPA's Hybrid Forecasting Competition (HFC), in their own words:
The HFC program seeks to develop and test hybrid geopolitical forecasting systems. These systems will integrate human and machine forecasting components to create maximally accurate, flexible, and scalable forecasting capabilities.
The idea essentially was to study how one can combine human forecasts from a Good Judgment-style platform with machine-generated forecasts to answer questions like. We were part of USC ISI's SAGE team. My specific contribution was an automated forecasting system that used data provided by a corresponding ISI data collection platform in combination with some parsing logic to “understand” a multinomial question in order to apply univariate time-series models and turn them into probabilistic forecasts for the answer options. About half of the questions overall were at least in principle amenable to a univariate time-series based solution, and of those the system covered 2/3rds, with an overall accuracy that for the optimal models approached the best combination of human forecasts.
- Years: 2017 - 2020
- Publications: IJCAI
- R, Docker, Python (Flask, but eventually R plumber instead)
Forecasting irregular leadership changes
Predicting IED and mine placements
- Years: 2013 - 2016
W-ICEWS is a global crisis monitoring and forecasting system (here's the Lockheed-Martin page on this). A lot of people are familiar with the ICEWS event data that was generated as part of this project, but there was actually a forecasting component for five different forms of crises as well. I joined Duke at the tail end of the project, when the forecasting component was mostly already in place, and helped a bit with maintaining and improving it. This included improvements to the R package that was used to deliver monthly forecast updates, as well as to the general monthly forecast update pipeline. I also setup and maintained a MySQL server for lab use to host the raw ICEWS event data, which consists of 70+ database tables totalling 200GB at the time.
- Years: 2012 - 2014
- Comparing GDELT and ICEWS Event Data
- spduration: split-population duration models for R; a package I worked on during this time as well.