PARKED · compute
Hypothesis
SSA decomposes a series into trend, oscillatory, and noise components via the eigentriples of its trajectory matrix; the reconstructed trend is a lower-lag filter than moving averages.
Math — trajectory matrix SVD
Embed the series into an $L\times K$ trajectory matrix and SVD it:
$$ X = \sum_{i} \sqrt{\lambda_i}\,U_i V_i^\top $$
Group leading eigentriples as "trend", reconstruct by diagonal averaging.
Method
Causal SSA on a sliding window; the open question is whether grouped reconstruction lags less than an EMA of equal smoothness, net of recompute cost.
Unlike N-034/035, SSA can be made strictly causal. Parked because per-bar SVD on a long window is expensive and the lag advantage over a tuned EMA looked marginal in spot checks. Revisit if a cheap incremental SVD is wired.
Causal is necessary but not sufficient — a filter also has to beat the cheap baseline (an EMA) by enough to justify its cost.