Chan Compass · US Markets
Multi-timeframe 缠论 (Chan theory) signal engine — monthly → weekly → daily → 60m → 30m → 15m → 5m nested-interval confirmation — plus sector capital rotation and a fully local llama.cpp research brain.
Not run yet — press “Run now” or wait for the 18:10 ET schedule.
Tomorrow's plan — long-hold mode (sorted: BUY → SELL → HOLD → WAIT)
Decision log — the engine's full multi-timeframe ruling chain (engine output is in Chinese; use the button for an English explanation).
Capital rotation across the 11 SPDR sector ETFs (full S&P 500 coverage). Flow proxy = price change % × dollar volume; RS = return minus SPY. True per-sector fund-flow feeds are paid data — this is the standard free proxy.
Press refresh (or Run analysis on the Signals tab).
1-Day (today's rotation)
5-Day (week trend)
20-Day (month trend)
Daily rule: for each holding, only today's news is checked. News found → AI brief is pushed below. No news → the ticker is ignored (listed under Quiet today).
Multi-step research agent (fully local): PLAN → 5 evidence tools (fundamentals · quarterly financials · price action · the Chan engine itself · news) → section-by-section analysis → report. Every step is logged to a JSON agent trace on persistent storage. New tickers entering the signal pool get a report auto-generated by the daily pipeline.
Report library (auto + manual, stored on /data):
✅ Persistent storage bucket mounted at /data — data cache, model files, reports, traces and the llama.cpp runtime all survive restarts.
Schedule: Mon–Fri 18:10 America/New_York (market closes 16:00 ET; by 18:10 the official daily bar has settled — that's 07:10 next morning Beijing time).
Now (ET): 2026-06-12 06:03 · Last run: never
⚠️ On free Space hardware the app sleeps when idle and the timer can't fire; use Run now, or upgrade to always-on hardware for unattended updates.
Saved agent traces (/data/traces):
SPCX_20260612-081714.jsonSPCX_20260612-075731.jsonNVDA_20260612-041216.jsonNVDA_20260612-035612.jsonJPM_20260612-041418.jsonGOOGL_20260612-041349.jsonAMZN_20260612-041318.jsonAMD_20260612-041256.jsonAAPL_20260612-041225.json
📡 To claim the open-trace badge: download this folder and push it to the Hub as a dataset (huggingface-cli upload).
All AI runs locally through llama.cpp (llama-cpp-python) with Qwen3 GGUF weights — every option is far below the 32B-parameter cap, and nothing leaves the machine. First load installs the llama.cpp runtime + downloads the GGUF (one-time, usually 1–3 min; worst case ~15 min if it has to compile). Signals/rotation/news never depend on it.
Sub-agent pool: fast Translator/Narrator (Qwen3-1.7B, fixed) handles Explain / rotation narrative / news briefs; deep Analyst writes research reports. Each has its own lock — they run in parallel. Pick the Analyst model below:
⚪ Translator sub-agent (Signals · Explain) — not loaded yet (auto-loads at startup)
⚪ Narrator sub-agent (Sector Rotation) — not loaded yet (auto-loads at startup)
⚪ Reporter sub-agent (News · Research support) — not loaded yet (auto-loads at startup)
⚪ Analyst sub-agent (Auto Research) — not loaded yet (auto-loads at startup)
Each sub-agent has its own lock — Explain / narrative / research run in parallel without “model busy”.
Chan Compass · educational tool, not investment advice · data: Yahoo Finance · design language: Adobe Spectrum 2