SSM-JTK – Conservative fit & model selection (1.4)

Policy. Admit new terms only if they earn their place.

Training data.

  • calendar grid: daily or sparse (e.g., D1/D15), fixed time base
  • unwrap rule: carry +360 so L_unwrapped(t) is continuous

OLS targets/design.

  • Fixed-n (Moon/outer):
    y = L_unwrapped - n*t ; X = [1, sin(w*t), cos(w*t), ...]
  • Free-n (Sun/inner):
    y = L_unwrapped ; X = [1, t, sin(w*t), cos(w*t), ...]

Model score (BIC).
BIC = k*log(N) + N*log( max(RSS/N, 1e-16) )

Event-aware loss (degree still dominant).
loss = 1.0*MAE_deg + 0.3*cusp_MAE + 0.4*MAE_speed

Gate (strict).
Admit the extra term iff (BIC_extra <= BIC_base - 6) and (loss_extra < loss_base).

Selection guarantees.

  • Parsimony: base wins unless an extra is decisively better (DeltaBIC >= 6).
  • Event fidelity: crossings and stations influence selection, not just degree MAE.
  • Midpoint anchor: reduces phase coupling and edge bias.

Navigation
Back: SSM-JTK – Ephemeris-independent planetary kernel (1.3)
Next: SSM-JTK – Reproducibility & audit trail (1.5)