[wJubran]

What and why

Several features read files that only the old data server could produce, so they stayed empty in the open-source edition:

• the chart's BANDS overlay;
• the bands family in the Pulse Room;
• the whole REPLAY page.

backend/local_analysis/ now writes those files from the active feed's daily bars. This covers MetaTrader 5 and the simulator, and any future data source.

Bands (local_analysis/bands.py → {T}_bands.json)

• Anchor: each trading day's GBM / MJD / Bates forecast starts at that day's open, so the envelope follows the price.
• Width: taken from the last 60 trading days. The tails keep the cones' ~250-day shape: MJD's jumps, and Bates' long-run variance, vol of vol and correlation.
• Points: one history point per day, half a trading day after the open, plus a 20-day forward fan from the newest open. mean is the shared centre: every model has a zero expected log return, so it equals the anchor price.
• Format: the exact keys and formats the chart, the Pulse Room and the Replay read:
  • timestamps as "YYYY-MM-DD HH:MM" on the feed's clock;
  • bat_* for Bates;
  • _last_modified for the stale marker.

Replay page

• {T}_bands_replay.json (bands_replay_v2) holds every daily anchor's forward fan and its select point.
• {T}_historical_cones.json holds the past weekly and monthly cones.
• Quarterly cones now cover eight quarters and carry meta.anchors; the chart still reads the first four.
• Reuse: past anchors never change, so a daily rebuild reuses them. The first build of the replay files takes about 15 s per ticker in the background; after that it takes about 0.25 s.

Runner

• Each pass first writes the live files for every ticker (cones, quarterly, bands), so the overlays appear within seconds. It then writes the replay files.
• When the live files change, it rebuilds the Pulse Room snapshot.
• JSON is written with allow_nan=False.

Models

• Jump detection: jumps are now flagged relative to an EWMA of the volatility of their time. Before, one threshold covered the whole year. The new bands tests caught the problem: after a lasting rise in volatility, the new regime's ordinary days were all treated as jumps, and the recent volatility was underestimated. This also improves the MJD and Bates cones.
• MJD speed: quantiles are solved for all horizons at once. That is 5.6× faster, and results match the old code to 1e-15.

Also on this branch: two building blocks for the next PRs (regimes and options), with their own tests. Nothing uses them yet.

• a Gaussian HMM (local_analysis/hmm.py);
• option-chain arithmetic for GEX, the gamma flip, walls and max pain (local_analysis/options_math.py).

Docs

• The READMEs (EN/AR) and the Learning Center describe the bands and the Replay files.
• The README claimed that Money Flow fetches live data. It doesn't: the page reads cache files like the other panels. That claim is removed.
• The changelog is updated.

How it was tested

• python -m pytest passes in backend/: 164 tests on Python 3.10 and 3.13. New tests (tests/test_local_bands.py) cover:
  • the shape and ordering of the lines;
  • √τ width scaling;
  • recent-volatility estimation, and jumps judged against the volatility of their time;
  • no look-ahead;
  • the replay file matching the live history;
  • reuse of past anchors;
  • eight quarterly anchors;
  • the runner's files and the rule against overwriting another producer's files;
  • the routes (/api/bands, /api/historical_cones, /api/bands_replay, /api/quarterly_cones);
  • a Pulse Room snapshot with the bands family.
• UI changes: the Learning Center text only. Rebuilt with CI=true npm run build; frontend_build/ matches the source build.
• Tried it with python launcher.py --sim:
  • Chart: BANDS on an AUDUSD M15 chart, with the envelope following the price and a widening forward fan.
  • Replay: a 6-month XAUUSD replay on H4 with weekly cones and bands, 52 anchors loaded.

Screenshots

Checked in a local browser run; the screenshots are not attached here.

---

Generated by Claude Code