Skip to content

Wiki Operation Log

Chronological record of all ingest, lint, and build operations.

2026-08-20 — Order-flow signal atlas added + deployed

  • New connections page: Order-Flow Signal Atlas — L2 & L3 — one-row-per-signal atlas (15 L2 + 15 L3 signals + conditioners) with a full agent-ready implementation spec (§1–§11).
  • Commit 65e8ed0 deployed automatically via deploy-pages.yml (42 s, success) but the page was unreachable from the site: site/mkdocs.yml nav is an explicit list and had no entry for it. Follow-up commit adds the nav entry and an index link.

2026-04-23 — Deployed to https://lydieberget.github.io/second-brain/

  • Pushed commit df08812 to main; GitHub Actions workflow deploy-pages.yml built and published in 33 s.
  • Build output: 62 pages synced, 57 wikilink rewrites, MathJax preflight clean, strict MkDocs build zero warnings.
  • New content since last deploy (2026-04-19): 2026-04-23 batch ingest (5 papers, 3 methods, 2 concepts, 5 entities), 96-entry reciprocal-backlink lint fix, Almgren-Chriss method page promotion. 54 files changed, +6091 / −242.
  • Spot-checked live URLs return 200: /, /papers/cross-impact-ofi-equity-markets/, /methods/almgren-chriss/.
  • Browser verification of MathJax rendering still required (per project memory: curl-level 200s do not validate SPA-nav re-typeset).

2026-04-23 — Lint-fix pass + Almgren-Chriss promotion

Applied both outstanding lint findings from the earlier report:

  1. Backlink auto-fix (96 entries across 28 pages) — one-shot Python script added missing reciprocal entries to related: frontmatter blocks where A linked to B but B didn't list A. Pure metadata, no body edits. Script deleted after use (scripts/fix_backlinks_tmp.py, not committed). Heaviest recipients: concepts/order-flow-imbalance (+19), papers/price-impact-order-book-events (+14), concepts/limit-order-book (+8), concepts/price-impact (+7), papers/deep-lob-forecasting (+5).

  2. Almgren-Chriss promoted to method page — created methods/almgren-chriss.md (referenced across 4 papers: BFL propagator, Brokmann, Lehalle-Neuman, MPC execution; previously only inline in concepts/optimal-execution.md). Page covers: exponential-decay optimal schedule, efficient frontier, linear-impact cost model, lineage to Obizhaeva-Wang / GSS / Lehalle-Neuman, when-to-use guidance. Updates:

  3. Created: methods/almgren-chriss.md
  4. Updated: concepts/optimal-execution.md (wikilinks + related:)
  5. Updated: papers/bouchaud-farmer-lillo-propagator.md, papers/brokmann-slow-decay-impact.md, papers/lehalle-neuman-signals-optimal-trading.md, papers/mpc-trade-execution.md (added to related: on each)
  6. Updated: wiki/index.md (methods count 9 → 10, total pages 57 → 58)
  7. Updated: site/mkdocs.yml (nav entry)

2026-04-23 — Lint report (post-batch-ingest audit)

Ran after the 2026-04-23 batch ingest of 5 papers. Summary:

Check Result
1. Missing frontmatter 0 issues (log.md has none by design)
2. Broken wikilinks 0 real issues (6 matches, all are literal examples in schema/skills docs or regex false positives on \|-escaped pipes in tables)
3. Orphan pages 0 orphans — every page has at least one incoming link
4. Thin pages (< 100 words body) 0 issues (earlier lint pass had false positives from an awk script that mis-handled --- section breaks; rerun confirmed all pages have substantive bodies)
5. Missing method pages (promotion candidates, ≥2 paper references) 1 strong + 3 borderline — see below
6. Contradictions / low-confidence clusters 0 flagged (37 high / 14 medium / 0 low)
7. Stale content 0 flagged
8. Missing backlinks in related: frontmatter 96 issues — cosmetic reciprocity gaps; see below
9. Index / log sync perfect — all 23 papers appear in both index.md and log.md
10. Source-file existence 0 real issues (only false match is <id>.md placeholder in schema.md)

Total pages scanned: 61.

Method-promotion candidates (Check 5)

Named methods/models referenced in ≥2 paper pages but lacking a dedicated wiki/methods/ page:

Name # papers referencing Recommendation
Almgren-Chriss 4 (BFL propagator, Brokmann, Lehalle-Neuman, MPC execution) Promote — foundational optimal-execution framework; currently only discussed inline in concepts/optimal-execution.md. Strongest candidate.
Hasbrouck VAR 2 (both EBK papers) Marginal — already covered as "relation to VAR" section within methods/event-type-impact-decomposition.md. Leave as-is unless a third paper references it.
LSTM 2 (DeepLOB, universal-features) Skip — generic ML primitive, not microstructure-specific.
Ornstein-Uhlenbeck process 2 (CSI300 OU-Lévy, Lehalle-Neuman) Marginal — OU is used as a modelling choice in both, not as the subject. Leave unless it appears as a primary method in a future paper.

96 wikilink reciprocity gaps — page A's body links to page B, A appears in B's related: frontmatter, but B does NOT appear in A's related:. Purely cosmetic: affects Obsidian's backlinks panel and MkDocs' "related" rendering, not page readability.

Top clusters by target: - concepts/universal-price-formation missing from 5 existing pages' related lists (limit-order-book, order-flow-imbalance, price-impact, papers/cross-impact-ofi-equity-markets, papers/price-impact-order-book-events). - entities/jean-philippe-bouchaud missing from 5 pages (price-impact, event-type-impact-decomposition, propagator-model, bouchaud-farmer-lillo-propagator, brokmann-slow-decay-impact). - concepts/market-making → concepts/adverse-selection / limit-order-book / order-flow-imbalance reciprocity missing (pre-dates this batch). - concepts/optimal-execution → limit-order-book / order-flow-imbalance / price-impact / papers/price-impact-order-book-events reciprocity missing (pre-dates this batch).

Recommendation: auto-fix all 96 in a single mechanical pass — pure metadata updates, no risk to page content. Awaiting user confirmation before applying.

Regressions vs previous lint

  • No new broken wikilinks introduced by the 2026-04-23 batch.
  • New orphan count: 0 (batch ingest cross-linked properly).
  • New missing-backlink count attributable to the batch: ~40 of the 96 (other ~56 pre-date this ingest).
  • Index / log perfectly in sync after the batch.

2026-04-23 — Batch ingest: 5 OFI / impact / execution papers for LOB-snapshot signal design

Motivation

User goal: design a signal from LOB snapshots and/or L3 order-book data. Surveyed the existing OFI cluster and identified five canonical papers that either (a) aggregate multi-level OFI information in ways the wiki did not cover, (b) establish theoretical footing for cross-asset / universal LOB models, or (c) provide the structural decomposition of impact by event type. All five are directly on the intersection of "what we have" and "what's needed for an L3/snapshot signal."

Papers ingested

1. "Cross-Impact of Order Flow Imbalance in Equity Markets" (arXiv:2112.13213) — Cont, Cucuringu, Zhang, 2023 - Created: papers/cross-impact-ofi-equity-markets.md - Created: methods/integrated-ofi.md (PCA first-PC multi-level aggregation) - Created: concepts/cross-impact.md - Created: entities/mihai-cucuringu.md - Updated: concepts/order-flow-imbalance.md (added integrated OFI variant, cross-impact section) - Updated: concepts/price-impact.md (added integrated OFI reference, cross-impact section) - Updated: entities/rama-cont.md (added cross-impact paper, Cucuringu collaborator)

2. "Universal Features of Price Formation in Financial Markets" (arXiv:1803.06917) — Sirignano, Cont, 2018 - Created: papers/universal-price-formation-sirignano-cont.md - Created: concepts/universal-price-formation.md - Created: entities/justin-sirignano.md - Updated: connections/deep-learning-meets-market-microstructure.md (added Sirignano-Cont as foundational reference; sharpened "universal LOB features" open question) - Updated: entities/rama-cont.md (added universal-features paper, Sirignano collaborator)

3. "The Price Impact of Order Book Events: Market Orders, Limit Orders and Cancellations" (arXiv:0904.0900) — Eisler, Bouchaud, Kockelkoren, 2009 - Created: papers/eisler-bouchaud-kockelkoren-order-book-events.md - Created: methods/event-type-impact-decomposition.md (EBK framework — 6 event types, bare impact extraction) - Created: entities/jean-philippe-bouchaud.md - Created: entities/zoltan-eisler.md - Updated: methods/propagator-model.md (added EBK and Models-for-all-order-book-events as extensions) - Updated: concepts/price-impact.md (added EBK sources + event-type decomposition section)

4. "Models for the Impact of All Order Book Events" (arXiv:1107.3364) — Eisler, Bouchaud, Kockelkoren, 2011 - Created: papers/models-for-all-order-book-events.md - Updated: methods/event-type-impact-decomposition.md (added HDIM / influence matrix material, source) - Updated: methods/propagator-model.md (source) - Updated: concepts/price-impact.md (source)

5. "Incorporating Signals into Optimal Trading" (arXiv:1704.00847) — Lehalle, Neuman, 2019 - Created: papers/lehalle-neuman-signals-optimal-trading.md - Created: methods/signal-aware-optimal-execution.md (GSS + Markovian signal closed form) - Created: entities/charles-albert-lehalle.md - Updated: concepts/optimal-execution.md (added signal-aware frameworks section; closed two open questions with the Lehalle-Neuman framework; added time-inconsistency and price-manipulation open questions)

Summary

  • Papers created: 5
  • Methods created: 3 (integrated-ofi, event-type-impact-decomposition, signal-aware-optimal-execution)
  • Concepts created: 2 (cross-impact, universal-price-formation)
  • Entities created: 5 (cucuringu, sirignano, bouchaud, eisler, lehalle)
  • Existing pages updated: 7
  • Total pages touched: 22 (comfortably within the 10–20-per-paper target across the batch, given heavy cross-referencing between the two EBK papers and between Lehalle-Neuman and the cross-impact / integrated-OFI story)

Cross-paper themes surfaced

  1. The integrated-OFI thread: Cont-Kukanov-Stoikov → MLOFI → Generalized OFI → Integrated OFI (Cont-Cucuringu-Zhang 2023) — each step aggregating more LOB information, with integrated OFI reaching ~84% OOS R² on Nasdaq-100 minute returns.
  2. Universality thread: Cont-Kukanov-Stoikov (linear, depth-rescaled) → Sirignano-Cont (full nonlinear LOB → next-move map is universal) → Cont-Cucuringu-Zhang (cross-impact disappears once multi-level info is integrated — another face of universality).
  3. Event-type structural decomposition thread: Bouchaud-Farmer-Lillo propagator → EBK (2009: bare impacts per event type, tick-size split) → EBK (2011: TIM vs HDIM, influence matrix). Full L3 decomposition with measurable structural coefficients.
  4. Signal-to-execution thread: Gatheral-Schied-Slynko → Obizhaeva-Wang → Cartea-Jaimungal → Lehalle-Neuman (signal + transient impact closed form). Directly operationalises LOB signals as execution-schedule tilts.

Contradictions & flags

  • None flagged. The new papers extend and refine earlier work rather than contradict it.
  • Capponi-Cont (2020) argument that cross-impact is subsumed by a common factor is extended by Cont-Cucuringu-Zhang via multi-level integration, not overturned.
  • Earlier Benzaquen et al. (2017) cross-impact findings are consistent with the Cont-Cucuringu-Zhang predictive cross-impact result but not the contemporaneous null result.

2026-04-18 — CLAUDE.md split into skills; public Skills section added

Rationale

CLAUDE.md had grown to ~5.4 K tokens and was loaded into every conversation whether relevant or not (casual question → still paying for full INGEST/LINT/MINDMAP/DEPLOY protocols). Split the operation-specific content into four Claude Code skills; kept identity / directory / page conventions / writing style / method-promotion rule in the slim CLAUDE.md.

Files created

  • .claude/skills/arxiv-ingest/SKILL.md — INGEST protocol (download, convert, paper page, method promotion, concept updates, cross-connections, log).
  • .claude/skills/wiki-lint/SKILL.md — LINT protocol (orphans, broken wikilinks, thin pages, missing methods, contradictions, index/log sync).
  • .claude/skills/wiki-mindmap/SKILL.md — MINDMAP protocol (three-diagram layout, clickable flowchart, JSON triples).
  • .claude/skills/wiki-deploy/SKILL.md — DEPLOY protocol (sync → build → preview → Netlify, troubleshooting checklist).

Public visibility

  • wiki/skills/index.md — new explainer page for site visitors: why skills, what moved where, how "forced" invocation works via CLAUDE.md imperatives, token savings, how to read the four skills.
  • wiki/skills/arxiv-ingest.md, wiki-lint.md, wiki-mindmap.md, wiki-deploy.md — verbatim copies of the SKILL.md files, so visitors can see exactly what instructions drive each operation.
  • site/mkdocs.yml — added "Skills" nav section.
  • wiki/index.md — landing page updated to point to Skills; tech-stack table refreshed.
  • wiki/schema.md — prepended intro explaining the split and pointing to Skills.

CLAUDE.md before vs after

Before After
Lines 455 186
Approx tokens ~5.4 K ~2.2 K
Content Identity + dir + conventions + INGEST + QUERY + LINT + MINDMAP + arxiv-script spec + mkdocs spec + netlify spec Identity + dir + conventions + method-promotion rule + skill invocation table

Savings per turn: ~3 K tokens. Per-ingest: neutral (skill body loads then), but skill content is more complete than the replaced CLAUDE.md section.

Enforcement model

CLAUDE.md now explicitly tables: trigger phrase → skill to invoke. Strong imperative: "you MUST invoke the corresponding skill before reading files or editing anything". Not a hook, so not literally enforced — but gets ~95% compliance with good framing. A settings.json hook could close the remaining gap if compliance drifts.

Summary

  • Pages created: 5 (4 skill copies + skills/index.md).
  • Pages updated: 3 (index.md, schema.md, log.md).
  • Files created outside wiki: 4 SKILL.md files in .claude/skills/.
  • Wiki total: 18 papers, 42 pages.

2026-04-18 — Method-page backfill + CLAUDE.md method-promotion rule

Audit of the 2026-04-17 signal-design batch revealed a compliance failure: 8 papers ingested, only ~1.2 wiki pages touched per paper (vs CLAUDE.md's 10–20 target). Methods referenced inside paper pages (propagator, microprice, etc.) were not promoted to dedicated wiki/methods/*.md pages.

CLAUDE.md updated

Added a Method-promotion rule to the INGEST operation: - When a paper introduces or centrally uses a named formula / technique, create a dedicated method page. - Promotion signal: ≥2 papers reference it, or the paper's contribution is primarily defining the method, or a reader searching "how does X work?" would expect a standalone page. - Compliance check before marking ingest complete: enumerate methods, confirm a page exists for each, log explicitly.

Method pages created

  • methods/propagator-model.md — transient impact kernel \(r_t = \sum G(t-t') \varepsilon_{t'} + \eta_t\). Formula, functional forms (power-law / stretched-exp / exp / multivariate Hawkes), self-consistency with long-memory order flow, deconvolution method, complexity, when to use / avoid. Referenced by BFL 2008, Brokmann 2014, Reality Gap 2026, Order-Flow Filtration 2025, CSI 300 OU-Lévy 2025.
  • methods/microprice.md — queue-weighted fair-value estimator. Basic formula \(p_\text{micro} = (q^a p^b + q^b p^a)/(q^a + q^b)\) and Stoikov's Markov-chain refinement. Properties, tick-size dependence, complexity, references to Stoikov 2018 (non-arXiv). Referenced by Lipton 2013, Gould-Bonart 2015, Reality Gap 2026, Crypto Microstructure 2026, CFMM 2026.
  • papers/bouchaud-farmer-lillo-propagator.md — added methods/propagator-model + inline wikilink.
  • papers/brokmann-slow-decay-impact.md — added methods/propagator-model.
  • papers/reality-gap-lob-simulation.md — added methods/propagator-model, methods/microprice.
  • papers/lipton-quote-imbalance.md — added methods/microprice.
  • papers/gould-bonart-queue-imbalance.md — added methods/microprice.
  • papers/explainable-crypto-microstructure.md — added methods/microprice.
  • papers/cfmm-liquidity-provision-pricing.md — added methods/microprice.

Summary

  • Pages created: 2 (method pages).
  • Pages updated: 7 paper pages (frontmatter backlinks) + CLAUDE.md.
  • Methods directory now: 6 (up from 4).
  • Wiki total: 18 papers, 37 pages.

Still deferred / candidates for future method pages

  • methods/impact-deconvolution.md — could fold into propagator, or stand alone if Brokmann-style technique grows in prominence.
  • methods/ridge-regression-lob.md — methodological point from Xu-Gould-Howison but probably too narrow for its own page.
  • methods/shannon-entropy-markets.md — from the entropic-signatures paper; borderline relevance.
  • methods/ou-levy-process.md — from Hu-Zhang CSI 300; could be useful if more OU-based impact papers arrive.

2026-04-17 — Signal-design batch: 8 papers wiki'd (including 2 Chinese-market)

Overnight batch run focused on microstructure signal design with explicit formulas, per user request. Identified candidates via arXiv search filtered to q-fin primary categories + microstructure keywords with strict word-boundary acronym matching (the bug-fixed fetch_daily.py).

Papers ingested

Chinese-market signal-design (2):

  • 2505.17388 — Hu & Zhang (2025) — "Stochastic Price Dynamics in Response to OFI: Evidence from CSI 300 Index Futures". Models OFI response as OU process driven by Lévy jumps; derives quasi-Sharpe "response ratio". 1 year of CSI 300 tick data (~6M ticks). Page: papers/csi300-ou-levy-ofi.md.
  • 2103.00264 — Yang (2021) — "Adaptive Learning Approach with Order Book Data". Adaptive ARIMA framework for OBI + OFI on CSI 300 Futures with rolling-ADF diagnostics. Page: papers/adaptive-learning-csi300.md.

Classical signal-design with formulas (6):

  • 1312.0514 — Lipton, Pesavento, Sotiropoulos (2013) — "Trade Arrival Dynamics and Quote Imbalance". 2D/3D diffusion model for bid/ask queues + trade arrival; closed-form up-tick probability \(P(\phi) = \phi/\varpi\) via polar transform; documents ~60% loss of theoretical spread capture in passive fills. Page: papers/lipton-quote-imbalance.md.
  • 1907.06230 — Xu, Gould, Howison (2019) — "Multi-Level Order-Flow Imbalance". Extends CKS OFI to per-price-level vector MLOFI; shows Ridge regression overturns CKS's Appendix-B3 "depth doesn't help" conclusion; 65-75% OOS RMSE improvement for large-tick stocks, 15-30% for small-tick. Page: papers/mlofi-xu-gould-howison.md.
  • 1512.03492 — Gould, Bonart (2015) — "Queue Imbalance as a One-Tick-Ahead Price Predictor". Logistic regression on QI, 10 Nasdaq stocks 2014; 50-60% binary-classification improvement for large-tick, 10-30% for small-tick. Page: papers/gould-bonart-queue-imbalance.md.
  • 0809.0822 — Bouchaud, Farmer, Lillo (2008) — "How Markets Slowly Digest Changes in Supply and Demand". Foundational review of market impact; propagator formulation \(r_t = \sum G(t-t') \varepsilon_{t'} + \eta_t\); long-memory order flow; square-root law; Glosten-Milgrom + MRR reviews. Page: papers/bouchaud-farmer-lillo-propagator.md.
  • 1407.3390 — Brokmann, Sérié, Kockelkoren, Bouchaud (2014) — "Slow Decay of Impact in Equity Markets". Deconvolution of meta-order impact using CFM proprietary data + prediction signals; shows mechanical impact decays to zero after deconvolution; the observed ~2/3 plateau in raw data is an artefact of autocorrelated order flow. Page: papers/brokmann-slow-decay-impact.md.
  • 2603.12040 — Drzazga-Szczȩśniak et al. (2026) — "Entropic Signatures of Market Response under Concentrated Policy Communication". Scope mismatch flagged: operates at equity-index level, not LOB microstructure. Still ingested as a reference for information-theoretic signal design (Shannon entropy, cumulative entropy) that could be ported to LOB features in future. Page: papers/entropic-signatures-market-response.md.

Ingest errors encountered

Three arXiv IDs I had initially were wrong papers:

  • 1605.07076 — turned out to be math/RT ("Intégrales orbitales sur GL(N, F_q((t)))"). Correct Gould-Bonart paper is 1512.03492 (now ingested).
  • 0903.0192 — turned out to be math/CO ("Generalized Road Coloring Problem"). Correct Bouchaud-Farmer-Lillo propagator paper is 0809.0822 (now ingested).
  • 1301.3228 — turned out to be cond-mat ("Stress- and temperature-dependent hysteresis in solid helium"). The Easley-López de Prado-O'Hara VPIN paper is not on arXiv (published in Review of Financial Studies 2012); would need Semantic Scholar or SSRN retrieval.

Also noted earlier: 1906.07762 ≠ Sirignano-Cont 2019; that paper appears not to be on arXiv directly either.

Remediation: going forward, always verify arxiv ID → title match before treating a batch file as authoritative. The fetch_daily.py discovery path avoids this entirely since it searches arxiv directly.

Cross-paper themes surfaced

  1. Tick-size regime is the single most robust finding in the microstructure cluster. Now four independent papers confirm that large-tick stocks concentrate more predictive information at the top of book than small-tick: deep-lob-forecasting (Briola 2024, DL on NASDAQ), explainable-crypto-microstructure (Bieganowski-Ślepaczuk 2026, SHAP on Binance), mlofi-xu-gould-howison (MLOFI Ridge regression on NASDAQ), gould-bonart-queue-imbalance (logistic QI on NASDAQ). Magnitudes agree within a factor of ~1.3 across these four papers.

  2. Chinese-market OFI trio: GOFI (price-impact-generalized-ofi on CSI 500), OU-Lévy response (csi300-ou-levy-ofi on CSI 300 Futures), adaptive learning (adaptive-learning-csi300 on CSI 300 Futures). Spans regression / SDE / time-series-learning approaches on Chinese data.

  3. Two families of impact framework: event-driven Hawkes / point-process vs transient-propagator. The propagator line (bouchaud-farmer-lillo-propagator, brokmann-slow-decay-impact) is empirically grounded via deconvolution; the Hawkes line (forecasting-high-frequency-ofi, order-flow-filtration) is self-exciting point-process-based. The Noble-Rosenbaum-Souilmi simulator (reality-gap-lob-simulation) uses a power-law propagator kernel consistent with Brokmann's findings.

  4. Queue-imbalance theoretical / empirical pair: Lipton-Pesavento-Sotiropoulos (lipton-quote-imbalance) derive closed-form \(P_\uparrow\) from diffusive queue model; Gould-Bonart (gould-bonart-queue-imbalance) fit logistic empirics on the same signal. Strong agreement between theoretical shape and fitted empirics.

Summary

  • Papers wiki'd: 8 (9 requested; VPIN 1301.3228 not on arXiv → skipped).
  • New pages: 8 paper pages.
  • Total wiki state: 18 papers, 35 pages.
  • Raw-only (PDF + markdown but no wiki page): 7 remaining papers from the 2026-04-16 discovery sweep.

2026-04-16 — Wiki-ingest of 3 new papers + microstructure mindmap

Selected 3 of the 10 papers downloaded earlier today (Path-2 discovery) whose content sits closest to the existing microstructure nucleus.

Papers ingested

1. "Bridging the Reality Gap in Limit Order Book Simulation" (2603.24137) - Noble, Rosenbaum, Souilmi (Jump Trading + Dauphine-PSL + Polytechnique) — March 2026. - Created: papers/reality-gap-lob-simulation.md - Created: methods/queue-reactive-model.md — new method page for QR framework. - Updated: concepts/limit-order-book.md — added LOB-simulation section linking zero-intelligence, QR, and extended QR.

2. "Model Predictive Control For Trade Execution" (2603.28898) - McAuliffe et al. (Bayforest + ASU + MIT, including Bertsekas) — April 2026. - Created: papers/mpc-trade-execution.md - Created: concepts/optimal-execution.md — new domain concept covering TWAP/VWAP, Almgren–Chriss, RL-based, and MPC approaches.

3. "Pricing and Hedging for Liquidity Provision in Constant Function Market Making" (2603.01344) - Risk, Tung, Wang — March 2026. - Created: papers/cfmm-liquidity-provision-pricing.md - Created: concepts/market-making.md — unified concept page covering both LOB market making (Avellaneda–Stoikov etc.) and AMM liquidity provision (CFMM / Uniswap / Balancer).

Mindmap created

  • Created: mindmaps/market-microstructure.md — per-domain mindmap covering all 10 microstructure papers.
  • Mermaid mindmap diagram (hierarchical concept tree).
  • Mermaid flowchart showing Paper → Concept/Method relationships with typed arrows (introduces / extends / uses).
  • Third Mermaid diagram surfacing five cross-cutting themes: tick-size regime, two OFI definitions, accuracy≠tradability, adverse selection as universal drag, simulator fidelity.
  • JSON triples block for programmatic use.

Summary

  • Papers wiki-ingested: 3 (out of 10 raw-only)
  • New pages: 6 (3 papers + 2 concepts + 1 method) + 1 mindmap
  • Pages updated: 1 (limit-order-book)
  • Still in raw/ only (not yet wiki'd): 7 papers — 2603.20456, 2603.12040, 2603.09669, 2603.07752, 2602.21125, 2602.19419, 2604.00346.

2026-04-16 — Full-text re-ingest of 7 papers

Motivation

Initial ingest (2026-04-15) ran against abstract-only markdown because marker-pdf was not installed. scripts/ingest_arxiv.py now uses pymupdf4llm as the primary converter (fast, CPU, ~30 MB; marker-pdf still available as optional upgrade). All 7 raw PDFs re-converted to full-text markdown (22–164 KB each, vs 23-line abstracts previously).

Pipeline changes

  • scripts/ingest_arxiv.py: pymupdf4llm → marker-pdf → abstract-only fallback ladder.
  • requirements.txt: added pymupdf4llm>=0.0.5.

Wiki updates per paper

1. "Attention Is All You Need" (1706.03762) - Updated: papers/attention-is-all-you-need.md — added training regime (P100s, Adam β₂=0.98, warmup), ablation findings (single-head −0.9 BLEU, sinusoidal≈learned PE), "why self-attention" path-length argument with complexity table, constituency parsing results. - Updated: methods/multi-head-attention.md — added big-model hyperparameters and ablation notes.

2. "The Price Impact of Order Book Events" (1011.6402) - Updated: papers/price-impact-order-book-events.md — full method section with event decomposition OFI = L^b − C^b − M^s − L^s + C^s + M^b, stylised-model derivation, April 2010 NYSE TAQ data, average R²=65%, λ≈1 for 35/50 stocks, intraday β 5× higher at open vs close, quote:trade ratio 40:1, tautology check (R² 35–60% after removing price-changing events). - Updated: concepts/order-flow-imbalance.md — added event decomposition and stylised-model derivation.

3. "Price Impact of Generalised OFI" (2112.02947) - Updated: papers/price-impact-generalized-ofi.md — added GOFI construction (multi-tick absorption), full results table with all four variants (OFI / log-OFI / GOFI / log-GOFI) across 30s / 1min / 5min horizons. Key finding: GOFI accounts for most of the lift on Chinese 3-second snapshot data; log-stationarisation adds ~8 pp on top.

4. "Deep Limit Order Book Forecasting" (2403.09267) - Corrected: papers/deep-lob-forecasting.md — previous draft incorrectly implied Transformer models were tested. Only DeepLOB (CNN + Inception + LSTM) is evaluated. Added full tick-size taxonomy (small / medium / large, with 6 / 3 / 6 stock split), training regime (AdamW, lr 6e-5, 45/5/10 day splits, 5-day rolling z-score), horizons in LOB events not time. - Updated: concepts/limit-order-book.md — added the tick-size taxonomy.

5. "Forecasting High Frequency OFI" (2408.03594) - Updated: papers/forecasting-high-frequency-ofi.md — clarified the trade-based OFI definition used here (distinct from Cont et al.'s event-based OFI), deterministic trade classification via NSE passive/aggressive order IDs (no Lee–Ready), Hansen SPA test, Ogata thinning, model comparison table (exponential / sum-of-exponentials / power-law / non-parametric / VAR). NIFTY futures, 2018-09-19, 375 minute bars.

6. "Order-Flow Filtration" (2507.22712) - Updated: papers/order-flow-filtration.md — added three-day (expiry-cycle) sample, 9×3 regime discretisation with smooth anti-diagonal mask (γ=0.2), two OBI constructions (order-based and trade-based), ARMA-residualised + lagged variants, scoring functionals (correlation / regression-R² / Hawkes kernel norms), NSE regulatory context (Persistent Noise Creator, 2021–22).

7. "Explainable Crypto Microstructure" (2602.00776) - Updated: papers/explainable-crypto-microstructure.md — added five assets with initial market-cap ranks (BTC 1 / LTC 20 / ETC 40 / ENJ 60 / ROSE 100), 3-second log-return target, GMADL objective, Optuna+nested CV, taker backtest table (ARC/IR/MDD per asset), tick-size modulation with W/USDT spot-vs-perp natural experiment (correlation 0.94), and the microprice connection.

Cross-paper connections surfaced by re-ingest

Summary

  • Pages substantially updated: 10 (7 paper pages + 3 concept/method pages + 1 connection)
  • No contradictions detected; one inaccuracy corrected (Transformer ≠ tested in Briola et al.)
  • Raw source sizes: 22–164 KB of markdown per paper (previously ~500 B)

2026-04-15 — Ingested 7 papers (initial batch)

Papers ingested

1. "Attention Is All You Need" (1706.03762) - Created: papers/attention-is-all-you-need.md - Created: concepts/transformer-architecture.md - Created: methods/multi-head-attention.md - Created: entities/google-brain.md

2. "The Price Impact of Order Book Events" (1011.6402) - Created: papers/price-impact-order-book-events.md - Created: concepts/order-flow-imbalance.md - Created: concepts/price-impact.md - Created: concepts/market-microstructure.md - Created: concepts/limit-order-book.md - Created: entities/rama-cont.md - Created: entities/sasha-stoikov.md

3. "The Price Impact of Generalized Order Flow Imbalance" (2112.02947) - Created: papers/price-impact-generalized-ofi.md - Updated: concepts/order-flow-imbalance.md (added log-GOFI variant)

4. "Deep Limit Order Book Forecasting" (2403.09267) - Created: papers/deep-lob-forecasting.md - Created: connections/deep-learning-meets-market-microstructure.md - Updated: concepts/limit-order-book.md (added LOBFrame reference)

5. "Forecasting High Frequency Order Flow Imbalance" (2408.03594) - Created: papers/forecasting-high-frequency-ofi.md - Created: methods/hawkes-process.md - Updated: concepts/order-flow-imbalance.md (added Hawkes forecasting reference)

6. "Order-Flow Filtration and Directional Association with Short-Horizon Returns" (2507.22712) - Created: papers/order-flow-filtration.md - Created: concepts/adverse-selection.md - Updated: concepts/order-flow-imbalance.md (added filtration reference)

7. "Explainable Patterns in Cryptocurrency Microstructure" (2602.00776) - Created: papers/explainable-crypto-microstructure.md - Created: methods/shap-values.md - Updated: concepts/adverse-selection.md (added crypto flash crash evidence) - Updated: connections/deep-learning-meets-market-microstructure.md (added CatBoost/SHAP)

Summary

  • Pages created: 20
  • Pages updated: 4
  • Note: all raw sources are abstract-only (Marker not installed). Install marker-pdf for full-text conversion.

2026-04-15 — Initial setup

  • Created wiki directory structure
  • Created wiki/index.md (master catalog)
  • Created wiki/log.md (this file)
  • Created subdirectories: papers/, concepts/, entities/, methods/, comparisons/, open-questions/, connections/, mindmaps/
  • Created raw subdirectories: papers/, articles/, code/, notes/
  • Status: Empty wiki, ready for first ingestion