01 / IMPLEMENTATION & CONTRIBUTION
What the work involves
Donald built raw-feature retrieval baselines, model-adaptation experiments, causal PCA, and persistence-head training around upstream time-series models. The related Kronos RAG workspace is a separate local study.
Technical depth
Feature fitting must respect time. Input and target units must match. A ranking metric can improve while the position behavior becomes worse.
Data and assumptions
Twenty-day stock windows, normalized raw-feature retrieval, and separate historical adaptation/ablation studies. Models including Chronos and Granite TTM are upstream.
Historical windows and labels need appropriate time boundaries. Architecture-level causal filters alone do not certify every evaluation as leakage-free.
The project family
chronos_knnA_Share/kronos_rag02 / THE EXPERIMENTAL RECORD
What the experiment taught
A trend-clarity objective made nearly every forecast positive. The simulator required negative trend clarity to exit, so its exit mechanism stopped firing.
What changed
The work moved from neural embeddings to a raw-feature baseline, and later explored adaptation and persistence classification. The later head is not claimed to fix the earlier failure.
03 / RESULTS
What came out of it
The research framework connects retrieval, forecasting, and decision-rule evaluation. Ablations exposed an objective that disabled simulated exits, while regime-selected retraining did not outperform its rolling baseline in the documented study.
The lesson in the work
Test the entire prediction-to-decision chain. A smooth output is not an economic objective.
04 / SUPPORTING EVIDENCE
Follow the source
Implementation notes, project records, and supporting artifacts.
Source context & project scope
Upstream foundation models are not Donald inventions. No live-performance claim. Architecture time filters alone do not certify full label-availability correctness; simulator leverage and calibration caveats preclude promoting headline returns. Later persistence heads are not proven to fix the earlier failure.
Donald-attributed implementation and experiment history. Foundation models and their research are credited as upstream work.
Twenty-day windows are represented by normalized raw features and retrieved through Qdrant, with versioned preprocessing and evaluation boundaries.
SOURCE · 2026-09-17The study did not establish an improvement from weekly regime-selected retraining over the rolling baseline.
REPORT · 2026-03March 2026 negative ablation: the exit mechanism stopped firing.
SOURCE · 2026-09-17Classifier on TTM decoder hidden states; March 10 Donald-attributed training/inference commits.
SOURCE · 2026-09-17Canonical snapback factor is mean reversion against stale Kronos envelopes, not fresh forecast following.
SOURCE · 2026-09-17