Skip to content

lasso: give fresh S&P downloads a decade of headroom #37

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
tschm merged 1 commit into from
Oct 14, 2026
Merged

lasso: give fresh S&P downloads a decade of headroom #37

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 5 additions & 4 deletions preprints/lasso/lasso.tex
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -533,7 +533,8 @@ \section*{Code and data availability}
The lockfile fixes the software, not the hardware. Residuals and gaps near $10^{-13}$ come
out with different last digits on another CPU or BLAS build, so a rerun will not match
the printed values digit for digit. The tests therefore check them against tolerances
between $10^{-12}$ and $10^{-9}$. Costs are given only as ratios, which carry over between
between $10^{-12}$ and $10^{-9}$ on the note's data, and up to $10^{-8}$ on fresh downloads.
Costs are given only as ratios, which carry over between
machines; \texttt{uv run python/make\_figures.py timings} measures them.

The test problem and the factor models of Table~\ref{tab:robust} are simulated inside
Expand All @@ -548,9 +549,9 @@ \section*{Code and data availability}
adjusted closing prices from 2021-06-01 to 2026-06-01, keeping names with no more than
5\% of days missing. The note used the snapshot with SHA-256 \texttt{b5faa522\ldots9937e}.
A later download differs slightly, because the constituents change and Yahoo revises
adjusted prices. On one, from October 2026, the path's KKT residuals stay below
$4\times10^{-12}$, but the QP baseline drifts to $3.4\times10^{-10}$ at one corner, so the
tests allow $10^{-9}$ for that gap on any download other than the note's snapshot.
adjusted prices. On two from October 2026 the path's KKT residuals stay below
$4\times10^{-12}$, but the QP baseline drifts by up to $9\times10^{-10}$ at some corners, so
the tests allow $10^{-8}$ for that gap on any download other than the note's snapshot.

\section*{Acknowledgements}

Expand Down
9 changes: 5 additions & 4 deletions preprints/lasso/python/make_figures.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1441,10 +1441,11 @@ def test_real_data() -> None:
"""Algorithm 1 on S&P 500 returns, through centred returns (every 10th corner)."""
snapshot = hashlib.sha256(SP500.read_bytes()).hexdigest() == SNAPSHOT_SHA256
rd = real_data(sp500_returns(), snapshot, every=10)
# The QP baseline matches the note's snapshot to 1e-10. A fresh fetch can be worse
# conditioned: on one from October 2026 the QP missed a corner by 3.4e-10 while the
# path's own KKT residual stayed below 4e-12, so the gap was the QP's.
assert rd.corner_gap <= (1e-10 if rd.snapshot else 1e-9), rd
# The QP baseline matches the note's snapshot to 1e-10. Fresh fetches are worse
# conditioned: on two from October 2026 the QP missed corners by up to 8.8e-10 while
# the path's own KKT residuals stayed below 4e-12, so the gap was the QP's. Allow a
# decade of headroom on data that are not the snapshot.
assert rd.corner_gap <= (1e-10 if rd.snapshot else 1e-8), rd
assert rd.minvar_gap <= 1e-12, rd
assert rd.kkt <= 1e-11, rd
# The warm-started baseline is exact to rounding on the note's snapshot. On a fresh
Expand Down
Loading