Align dy_d finite differences with updated R dy.d_

[OVVO-Financial]

Summary

Ports the revised R dy.d_ finite-difference estimator to Python.

R's dy.d_ was updated so its internal fd.estimates reduces every predictor row to its mean (xstar), duplicates that single column (the increasing-dimensions trick), and fits NNS.stack(method = 1) on the one-dimensional projection:

xstar.train.design <- cbind(Xstar = xstar.train, Xstar2 = xstar.train)
fd.estimates <- function(test_points) {
  xs <- xstar(test_points)
  NNS.stack(IVs.train = xstar.train.design, DV.train = y,
            IVs.test = cbind(Xstar = xs, Xstar2 = xs),
            method = 1, status = FALSE, order = NULL, folds = 1, ncores = 1)$stack
}

The Python port previously fed the full factor-expanded matrix to nns_stack(method=(1, 2), dim_red_method="equal"), which no longer matches.

Change

Rewrite _dy_d_stack_estimates to mirror the new fd.estimates exactly — row-mean reduction of both train and test points into a duplicated column, then nns_stack(method=1). All three finite-difference call sites (matrix band, vector band, and mixed derivatives) route through this single helper, so the entire estimator now follows the revised R definition. Every other part of the port (eval-point modes, bandwidth grid, zz dependence weighting, per-bandwidth weighted averaging, mixed-derivative stencil) already matched the new R and is unchanged.

Verification

• dy_d runs across all eval-point modes (mean, median, last, obs, apd, numeric scalar/vector) plus mixed=True, producing finite estimates consistent with the xstar projection.
• Confirmed the shift is intentional: the old mechanism reproduces the current tests/_r_cache.json baselines (7.56 ≈ cached 7.55; 0.0 = cached 0.0), so those fixtures reflect the old dy.d_.
• 11 dy invariant tests pass; ruff and mypy are clean.

Note on parity fixtures

The 8 dy_d parity cases in tests/parity/test_diff.py compare against tests/_r_cache.json, whose dy.d.scalar entries were generated from the old dy.d_. They stay red until the R cache is regenerated from the updated R package (triggered by the accompanying R update); the cases reconcile once the refreshed cache lands. No fixture was hand-edited from Python output.

🤖 Generated with Claude Code

---

Generated by Claude Code