docs: add a paper on the signals and the position engine

[tschm]

Adds docs/paper/main.tex (+ references.bib), giving the github-paper workflow
something to compile. 8 pages, builds clean with latexmk -pdf.

The derivation. osc() divides the EWMA difference by
s = sqrt(1/(1-f²) - 2/(1-fg) + 1/(1-g²)), asserted in the docstring without proof.
The paper proves it: the difference filter's coefficients sum to zero, so it kills the
unit root and the cumulative coefficients telescope to g^(m+1) - f^(m+1), whose sum
of squares is exactly s². So s is the standard deviation of the numerator under a
driftless unit-variance random walk, and the oscillator is a t-statistic of trend.
Simulation over 2e5 steps confirms it: realised sd within 1.4% of 1.0 for every
(fast, slow) pair tested, while s itself ranges over 1.09–6.13.

The empirical finding. Running the full engine over the 49-asset,
53-year panel in tests/tinycta/resources/prices_hashed.csv:

• shrink=1.0 fails outright — SingularMatrixError out of Engine.cash_position.
  The final EWMA correlation matrix over the 47 finite assets has condition number
  2.79e18 and smallest eigenvalue 9.16e-17, i.e. numerically rank-deficient even
  though the window (129) is longer than the cross-section (49). Shrinkage is
  load-bearing infrastructure here, not a tuning knob.
• The failure is preceded by leverage, not by a worse Sharpe. Mean gross exposure
  grows 11× from λ=0 to λ=0.99 and max gross grows 311× (706 → 219,768), while Sharpe
  stays inside [0.206, 0.327] and non-monotone. Someone watching only the Sharpe ratio
  would not see the ill-conditioning arriving.
• The parameter reads backwards. shrink2id(M, λ) = λM + (1-λ)I, so shrink=1.0
  means no shrinkage and shrink=0.0 discards correlations entirely — and 1.0 is the
  value that fails. Worth a docstring note or a rename, separately from this PR.

All Sharpe figures are gross of costs, in-sample, one signal, one panel; the paper says
so explicitly and draws no performance conclusion. Also documented: the profit-variance
feedback (κ=0.99, v₀=1, both hard-coded) leaves the book's absolute scale determined by
v₀ and the price series' units, and the per-timestamp correlation dict is O(T·n²) —
about 265 MB on this panel.

Code and docs are untouched — this PR only adds the paper.

[coderabbitai]

Warning

Review limit reached

@tschm, you've reached your PR review limit, so we couldn't start this review.

Next review available in: 11 minutes

Limit details: You’ve used the included review currently available.

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

How can I continue?

Wait for the limit to reset, then comment @coderabbitai review or push new commits to the PR.

An organization admin can change what happens after included review limits in Billing.

How do review limits work?

CodeRabbit enforces per-developer PR review limits within each organization.

For paid Pro and Pro+ reviews, CodeRabbit uses a developer's included PR review attempts over the past 7 days to set the current hourly allowance. At typical activity levels, the full plan allowance applies. Higher sustained activity can lower the allowance until earlier attempts leave the 7-day window.

Please refer docs for additional details.

Review details

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 7b54d7ad-66e3-4b9f-83db-0c0b1466e4cb

📥 Commits

Reviewing files that changed from the base of the PR and between c619ce8 and 2601395.

📒 Files selected for processing (2)

• docs/paper/main.tex
• docs/paper/references.bib

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.