TARPA: Retrieval Priors for Time-Series Foundation Models on Audience Retention

Frozen time-series foundation models (TS-FMs) promise off-the-shelf representations for forecasting tasks with little supervision. TS-FMs are pretrained to forecast future values from past observations. It remains unclear whether this knowledge transfers to predicting retention curves from multimodal content in a few-shot setting. We study this setting on a multimodal structured-output regression task: audience retention prediction on YouTube, with audio, visual, and text-derived per-second channels alongside tabular video-level metadata. We find two paired results — naive featurising fails, prior-conditioning succeeds: (1) A naive frozen TS-FM that consumes the multimodal per-second features does not improve over a no-FM baseline: the TS-FM was pretrained on raw univariate time series, not on engineered features. (2) Conditioning the same TS-FM on a non-parametric retention prior significantly improves how closely the predicted curve matches the true retention shape. Our method, TARPA (Template-Augmented Retrieval Prior Adapter), feeds the retrieval prior as an additional input channel to the frozen TTM backbone, giving the FM at least one channel in the format it was pretrained on. TARPA achieves the best RMSE and Spike-RMSE across all evaluated methods — a tree-based CatBoost regressor, FM-based predictors (TTM, Chronos-Bolt), and FM-free predictors. It matches the absolute-level metrics of the strongest non-FM baseline, and an ablation shows TTM is what drives this improvement. On this multimodal structured task, TS-FMs yield measurable gains when at least one input channel matches their pretraining format.