Debugging · Real anchors, not synthetic
"Sounds like we need to debug this... plot the position based on our oscillatory equation of motion with dampening, and visually see how it matched up with price." Each panel: the same real anchor point, real price forward vs. the model's projected price vs. a flat (no-movement) baseline — plus the model's own raw projected-return sequence underneath, with a zero line, so it's visible exactly what the model is committing to at each step.
theta = 2π/period, period 85–115 candles, borrowed from the SLOW
volatility envelope) was 24–32× too slow for how fast a real return's own sign actually flips,
keeping the projected return on one side of zero for 20+ consecutive steps. Fixed: theta
is now measured directly from the real (velocity, acceleration) phase's own candle-to-candle rotation
(1.75–1.81 rad/candle, measured, not assumed), plus a small acceleration-scale correction. Price MAE
collapsed from 7–10× worse than flat down to matching flat almost exactly (see the Primitives
Registry for the full numbers). The dashed gold line marks a "short horizon" (10 candles) — judge fidelity
there first: it's real but uneven across anchors (2%–7% average error in the first 10 candles here),
which is itself consistent with direction having no reliable predictable structure, not a modeling failure
specific to these four windows. 1 Minute panels below add a second (teal, dashed) line:
the model refit from ONLY the preceding 2 real "volatility cycles" (~2 days) instead of the whole history
— visually, it tracks the global fit almost exactly, matching the aggregate finding (400 real anchors)
that local refitting doesn't meaningfully beat the global fit or flat at 1m's own 15-candle horizon.