Extreme precipitation events are natural hazards that pose a significant risk to life and a commensurately large cost through property loss across multiple timescales. Unfortunately, extreme precipitation, especially extended-duration precipitation, remains as one of the most challenging hazards to forecast. The subseasonal to seasonal timescale, defined as the period between 2 weeks and 3 months, is characterized with low skill and a subsequent lack of informative forecast products for end-users. The present work constitutes a major step into the identification, characterization, prediction, and communication of subseasonal extreme precipitation. To identify an extreme precipitation period, thresholds for both total precipitation and the duration of the precipitation are used to identify events with sufficient length to accentuate the synoptic and subseasonal contributions. I then generate databases of 14-day extreme precipitation periods over the contiguous United States (CONUS) using observational datasets, reanalysis products, and climate model simulations, displaying the flexibility of my developed scheme. These datasets are used to quantify trends and synoptic-scale characteristics in subseasonal extreme precipitation. At present, the observational database is fully accessible in tabular format along with informational fact sheets describing the definition and climatology all available online. Rossby wave trains are shown to be present and co-located with anomalously strong moisture in several regions during both observed and simulated extreme periods. The observational database is also used to fit random forests and convolutional neural networks to understand the predictability of these extreme periods. The random forests show some skill above climatology in differentiating extreme and non-extreme days in three different regions of the U.S. with better predictability over the West Coast compared to the Central Great Plains. Although the critical success index (CSI) scores were relatively low, peaking at 0.23 along the West Coast, the random forests appear more skillful when using skill scores tailored towards extreme events. The convolutional neural networks produce probabilities of subseasonal extreme precipitation onto a map over the CONUS are verified using an object-oriented scheme (e.g., centroid offset, area ratio, etc.) and CSI scores reach 0.3. Although the skill is seemingly low, the networks are often identifying "close hits" and may still bring value to end users who are more tolerant to false alarms. The object-oriented scheme allows for flexibility in matching forecasts and observations that end-users can tolerate and gives a better indication of forecast usability. Although much work lies ahead to develop an operational forecast product, the work herein lays the foundation for future works to continually make advances into the S2S timescale.