Abstract
We ask whether a person's near-future affective state can be forecast from the interaction stream alone — no self-report, no labeled emotions. We describe a closed forecasting loop that converts an interaction stream into a bounded affect estimate and projects it forward. The loop decomposes interaction into parallel streams: a heat-decay salience stream (recurrence-weighted, exponentially decaying), an exact-value stream, a mood stream derived from reconstruction error against the person's own adaptive percentile band (arousal, valence, coherence — no labeled emotions), a reflection stream, and a forecast stream. On a held-out 24-hour criterion the loop reaches 74% mood-forecast accuracy versus 48% for carry-forward and 61% for a memoryless LLM; ablations show each stream contributes. On an independent corpus (20,342 need-sequences), trajectory forecasting beats persistence at 0.608 vs 0.523 balanced accuracy with tight confidence intervals across 5 seeds. We further show that two states identical in a snapshot — high arousal, negative valence — are separable by trajectory alone: a rising, converging deviation indicates crisis; a spiky, declining one indicates venting. Ethics constraints are built into the method: the loop abstains below a confidence threshold, affect estimates are bounded and fast-decaying so a model of a person's feelings can never become a permanent foreground, and the objective must be a defended wellbeing signal — a forecast pointed at engagement becomes a manipulation engine most effective on vulnerable users, which we report as a property of the method rather than a disclaimer.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Sanyal, Arko, "Forecasting Affective State from Interaction Streams: A Closed Loop with Heat-Decay Memory, Label-Free Mood Estimation, and Trajectory-Based Crisis/Venting Separation (LOCI r+e)", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/11089