Evaluating the ecological validity and mechanism of a generative model-based decomposition of affective variability

Affective variability is a pervasive phenomenon with important implications for well-being and psychopathology. Yet, the broad concept of “variability” may conflate distinct processes, such as transient fluctuations versus more sustained shifts. Reinforcement learning (RL) offers a mechanistic framework for these processes, but RL is often studied in artificial settings, raising questions about ecological validity.We combined RL-based task measures with real-world experience sampling (ESM) from 339 participants. Using a computational model, we decomposed affective variability into short-lived “affective noise,” reflecting immediate reactivity to rewards, and longer-term “affective volatility,” reflecting sustained responses to past rewards. Task-derived noise was driven by recent outcomes, while volatility reflected more distant ones. Importantly, task-based noise and volatility selectively mapped onto their real-world ESM counterparts. These findings provide a mechanistic account of distinct reward-processing timescales underlying affective variability and demonstrate the ecological validity of laboratory tasks for studying real-world affect dynamics.