Companion code and data artifacts for the research report
“Convex Core and the DNSR Model Suite: A Deterministic ETF Allocation Framework
for Drawdown-Controlled Compounding” (paper/).
Principal: David Reeves, DNSR Investments, LLC — https://www.linkedin.com/in/david-reeves-8a664524 Research system: the DNSR Agentic AI pipeline (with Anthropic Fable 5 and Opus 4.8), directed and reviewed by the principal, who is responsible for its use.
Not investment advice. All performance herein is hypothetical, backtested, net of modeled costs. Past performance does not guarantee future results. See the paper’s §15 (Limitations) and Disclaimer.
Two self-contained, visual explainers (inline-SVG charts, zero dependencies) that build an intuitive understanding of the models before the formal paper:
- Convex Core — an intuitive tutorial — what convex
means, how it lives inside
w_equity + w_convexity + w_duration = 1, the specific ETFs, the regime engine, a 20-year growth-of-$1 chart vs SPY and 60/40, and the equity-core options (SPY / VTI / VXUS / IWM, CAPE / valuation). - Convex Prime — the leveraged sibling — what leverage does to a convex payoff, how it’s sized / capped / sentiment-gated, and the honest risk/reward trade.
- A derivation from first principles (HTML) —
how the equation
w_equity + w_convexity + w_duration = 1and the volatility brake follow from maximizing expected log-growth, with each step marked rigorous or design choice (companion to §4.1). - The DIY edition — a plain-English guide (PDF) — the everyday-investor, no-math version: the one idea (avoid deep drawdowns), the three sleeves and illustrative funds, taxable-vs-IRA placement, and an optional volatility brake you can run in a spreadsheet. Education only — not investment advice.
- The Study That Predicted Its Own Conclusions — an essay (PDF) — third in the essay series: two pre-registered studies put deployment timing and destination choice on trial — when lost to deploy-on-receipt, where split into an answerable asset-class question, an unrankable fund question, and a century-scale factor verdict. On writing the honest headline in advance. Education only — not investment advice.
- Investment First Principles — the framework, with allocations (PDF) —
the bridge tier: derives the three-sleeve portfolio from the wealth-growth equation
g ≈ μ − ½σ², is explicit about what the equation does and does not settle (the convexity sleeve needs a left-tail/regime argument beyond mean-variance), then maps it to investor-profile allocations (accumulator → retiree), sequence-of-returns risk, and operating rules. Renders math inline on GitHub. Education only — not investment advice.
GitHub serves
.htmlas source. To read them rendered, open the files in any browser, or use the htmlpreview proxy — ▶ Core · ▶ Prime · ▶ Derivation. Educational illustrations of the model — not investment advice.
See exactly how Convex Core changed week by week — every ETF, its weight, the regime, the volatility brake, and growth-of-$1 vs SPY — across the entire backtest. The convexity sleeve, for instance, is empty in 2008 (those funds didn't exist yet — their weight folds into Treasuries), appears as BTAL by 2019, and is the full DBMF/KMLM/BTAL set by 2026. Point-in-time TARGET weights, suggest-only.
Each ledger opens to the current year for a fast first paint; pick an earlier start year or Full history to extend it (Convex Core/Prime back to 2006, RACE to 2007).
- Convex Core ledger — equity / convexity / duration sleeves + the volatility brake.
- Convex Prime ledger — the leveraged sibling, same three sleeves.
- RACE ledger — the seven-sleeve Regime-Adaptive Capital Engine (US-core / factor / intl / fixed-income / real-assets / crisis-alpha / cash).
Rendered (GitHub Pages): ▶ Convex Core · ▶ Convex Prime · ▶ RACE.
The paper’s central claim is methodological: a deterministic, auditable, reproducible allocation model, validated with pre-registered tests and published negative results. This repository exists so a third party can inspect and re-run the method rather than take the numbers on faith. It contains the model engine, the benchmark/report code, the computed result artifacts, and the tests — everything except the licensed vendor data (see Data).
| Path | Contents |
|---|---|
paper/ |
The report (PDF + self-contained HTML) and the Statistical-Robustness Addendum — Deflated Sharpe Ratio & Probability of Backtest Overfitting (PDF + HTML). |
tutorial/ |
Visual, intuitive tutorials — convex_tutorial.html (Convex Core) and prime_tutorial.html (Convex Prime). Self-contained HTML with inline-SVG figures. See Tutorials above. |
engine/tradeclassifier/ |
The deterministic model engine — convex_core.py (the flagship) and the shared point-in-time alpha/beta/regime panel it depends on (alpha_backtest, alpha, beta, features, optimizer, loaders, objective, regime, config). Zero fitted parameters; no LLM in the allocation math. |
config/ |
Model + regime configuration (regime_rules.yaml, ira_profile.yaml, etc.). The regime clause set is reproduced verbatim in the paper’s Appendix I. |
report/ |
pub_benchmarks.py (benchmark construction + metrics + bootstrap significance), build_pub_report.py (assembles the paper), run_tier2.py (sleeve ablations + crisis attribution), run_regime_wf.py (regime walk-forward validation), run_dsr_pbo.py (deflated-Sharpe + backtest-overfitting diagnostics), run_attribution.py (Shapley mechanism attribution), run_robustness.py (assumption breach frontier), build_addendum.py (assembles the addendum). |
results/ |
Computed model outputs (not raw vendor data): model_curves.json (growth-of-$1 per model), benchmarks.json, tier2.json, regime_wf.json, dsr_pbo.json, attribution.json (+ ATTRIBUTION.md — which mechanisms carry the record), robustness.json (+ ROBUSTNESS.md — how much assumption error the record survives), reproducibility_manifest.yaml. These reproduce the paper’s tables/figures (and the addendum) directly. |
tests/ |
test_convex_core.py — engine pins (vol brake, PIT fold, determinism, attribution non-invasiveness). |
The model is built from a local warehouse of daily ETF OHLCV and macro series sourced
from EODHD, FMP, and FRED. That data is licensed and is not redistributed here.
To rebuild it from your own vendor keys, see data/README.md for the
schema and the series list. The results/*.json artifacts let you reproduce every table
and figure in the paper without the raw data; the full engine run requires rebuilding
the warehouse.
pip install -r requirements.txt
# (1) Regenerate the paper (all tables + figures) from the shipped result
# artifacts — NO vendor data required. Self-resolves against results/.
# Verified: reproduces the distributed paper. Output:
# paper/DNSR_Convex_Core_Publication_REGENERATED.html
python report/build_pub_report.py
# (1b) Regenerate the Statistical-Robustness Addendum (DSR & PBO) from
# results/dsr_pbo.json — also NO vendor data required. Output:
# paper/DNSR_Convex_Core_Addendum_DSR_PBO_REGENERATED.html (+ .pdf)
python report/build_addendum.py
# (2) Run the engine unit tests (synthetic data, no warehouse needed):
PYTHONPATH=engine pytest tests/
# (3) Full engine re-run from scratch — REQUIRES a rebuilt data/etf_data.db
# (see data/README.md). These scripts read paths via the PAPER_ROOT env var:
export PAPER_ROOT=$(pwd)
python report/pub_benchmarks.py # rebuilds results/benchmarks.json (+ bootstrap)
python report/run_tier2.py # sleeve ablations + crisis attribution
python report/run_regime_wf.py # regime walk-forward validation
python report/run_dsr_pbo.py # deflated Sharpe + backtest-overfitting (DSR/PBO)Reproducibility status. Step (1) is self-contained and has been verified to regenerate the paper from the shipped
results/*.jsonartifacts. Steps (3) reproduce those artifacts from a rebuilt warehouse; they were extracted from a larger private system and read paths viaPAPER_ROOT. The PDF is produced from the HTML with a headless browser (e.g. Playwright/Chromium).
This work claims no novel strategy or anomaly. Convex Core is a synthesis of well-published premia — volatility-managed equity (Moreira & Muir 2017), crisis-alpha / trend-following (Hurst, Ooi & Pedersen 2017), and defensive / betting-against-beta (Frazzini & Pedersen 2014). The negative results corroborate the anomaly-replication literature (Hou, Xue & Zhang 2020; McLean & Pontiff 2016). The contribution is integrative and methodological: negative-results discipline, end-to-end determinism and reproducibility, significance-tested and honestly-calibrated claims, and an AI-conducted, pre-registered research process. See the paper’s §3.1.
Reeves, D. (DNSR Investments, LLC), with the DNSR Agentic AI system (Anthropic Fable 5 / Opus 4.8). Convex Core and the DNSR Model Suite: A Deterministic ETF Allocation Framework for Drawdown-Controlled Compounding. 2026.
Machine-readable metadata is in CITATION.cff — GitHub renders a
“Cite this repository” button from it (APA/BibTeX export).
- Code (engine, report/research scripts, tests): PolyForm Noncommercial License 1.0.0 —
source-available and fully auditable; noncommercial use permitted, commercial use requires a
separate grant from the licensor. SPDX:
PolyForm-Noncommercial-1.0.0. Full text inLICENSE. - Paper & documentation (
paper/, this README,data/README.md): CC BY-NC 4.0 (https://creativecommons.org/licenses/by-nc/4.0/).
Copyright © 2026 DNSR Investments, LLC. Licensor: David Reeves (https://www.linkedin.com/in/david-reeves-8a664524).