COTP: Center-Out Temporal Prediction

This repository contains the experimental artifacts and results for Center-Out Temporal Prediction (COTP), a center-conditioned approach to temporal representation learning. The study investigates whether a model given only a window from the middle of a time series can learn a useful representation of the temporal context before and after that window. Across 12 experiments, the approach was tested on synthetic signals and four real-world datasets: UCI Air Quality, Appliances Energy Prediction, Individual Household Electric Power Consumption, and Metro Interstate Traffic Volume. Key findings include: 48/48 learned configurations outperformed persistence on the final multi-dataset evaluation. COTP achieved approximately 57.4% mean improvement over persistence in the temporal baseline comparison. The frozen COTP latent representation outperformed a raw center-window representation for linear reconstruction in 24/24 conditions, with +11.91% mean improvement. The representation transferred across prediction horizons in 24/24 conditions, with +10.33% mean improvement. Variability classification improved in 19/24 conditions, with +1.19% mean improvement. However, COTP was essentially tied with a direct MLP (+0.10% mean improvement), so general architectural superiority is not claimed. DCT-based decoding also showed no consistent advantage across real-world datasets. The results support the feasibility of learning a compact, reusable temporal representation from a center window, while leaving the specific architectural advantage of COTP open for further investigation. The research was conducted with assistance from ChatGPT, Gemini, and Grok for brainstorming, experimental design, coding, debugging, analysis, and critical review. For transparency and reproducibility, this repository includes the Google Colab notebooks/code, experimental outputs, scorecards, predictions, and relevant LLM conversation records. Independent reproduction and testing are encouraged.