Log — Quant Trading Wiki

Append-only history. Entries: ## [YYYY-MM-DD] <op> | <title> (ingest/query/lint/split).

[2026-06-29] split | quant-trading-wiki created from _inbox `financial-ml` cluster (3 sources)

Spun out the moment a 3rd coherent source landed. The cluster had sat at two parked sources (kronos-financial-foundation-model, tensortrade); the TensorTrade park record explicitly flagged that “one more financial-ML / quant source → spin-out trigger (≥3).” banbot (arriving via Telegram) was that source. Founding corpus: 3 source pages + 5 Thing pages + 1 entity + synthesis.

kronos-financial-foundation-model (SoftwareSourceCode, T3, github.com/shiyu-coder) — first open-source financial-market foundation model: OHLCV tokenizer + decoder-only autoregressive Transformer; MIT open-weight, three sizes on HF. The predict layer. (Parked first in _inbox.)
tensortrade (SoftwareSourceCode, T3, github.com/tensortrade-org) — Python RL framework for trading agents (Gym-compatible component model; trains via Ray RLlib). The strategy/policy layer. (Parked second.)
banbot (SoftwareSourceCode, T3, github.com/banbox) — event-driven multi-symbol/strategy/ timeframe/account trading bot; one codebase for backtest + live; Go core; embedded Bayesian/TPE/ CMA-ES strategy optimization. The execution layer. (The trigger.)
Thing pages: algorithmic-trading (umbrella), backtesting, event-driven-trading, reinforcement-learning-trading, financial-time-series-foundation-model.
Entity: banbox (Organization) — banbot’s owner org. TensorTrade-org and Kronos’s author (shiyu-coder) entity nodes deferred (thin at founding).
Synthesis thesis: a trading system is a predict → decide → execute stack, bound by trustworthy, lookahead-free backtesting; the event-driven design is the shared structural guard and the basis of backtest/live parity. Open questions: do the three layers actually compose into one pipeline; all three sources are self-reported T3 (need a neutral benchmark); beyond crypto.
Scope boundaries recorded in CLAUDE.md: vs ../optimization-algorithms-wiki (optimizer as means, not subject — cma-es cross-linked), vs ../llm-providers-wiki (financial modality, not the text-LLM market — open-weight-models cross-linked), vs ../agentic-tooling-wiki (RL agents, not LLM agents). Registry block added to ../wikis.md; _inbox/kronos-financial-foundation-model.md and _inbox/tensortrade.md park records deleted (superseded by ingest).

[2026-06-30] ingest | Algorithmic trading (Wikipedia) — quality cycle floor-raise

First quality-cycle pass on the day-old spoke. The founding trio are all author-published T3 repos, so the floor was all-T3. Added algorithmic-trading-wikipedia (Article, T2, en.wikipedia.org) — the spoke’s first neutral third-party anchor. Grounds algorithmic-trading (field scale: ≈50% US equity / ≈80%+ FX volume algorithmic; the strategy taxonomy — arbitrage/pairs/mean-reversion/market-making/momentum/ event-arb) and backtesting (the backtest→forward→live validation ladder; forward-testing as the out-of-sample over-fit guard). Partly closed the synthesis “beyond crypto” open question — algo execution is already the norm in equities/FX, so the crypto founding bots are a recent entrant. Floor T3-only → T2 1 / T3 3. Publisher entity (Wikimedia) deferred per corpus precedent. Ran avoid-ai-writing (clean). +1 page (→10).

[2026-06-30] ingest | Strategy optimization — cross-spoke bridge to optimization-algorithms-wiki

User-directed: optimization-algorithms-wiki has pages directly reusable here (banbot tunes strategies with Bayesian/TPE/random/CMA-ES). Added bridge concept page strategy-optimization — subject = tuning a trading strategy (search space = strategy hyperparameters; objective = a backtested metric; per-eval cost = one whole backtest; the overfitting danger). It cross-links the algorithm pages as methods rather than re-deriving them, honoring the means-not-subject scope boundary. Maps banbot’s tuner menu onto the optimization spoke’s model-free (cma-es, metaheuristic-optimization) ↔ model-based (bayesian-optimization — TPE/Bayesian) axis, and borrows no-free-lunch-theorem + exploration-vs-exploitation. Threaded into banbot, algorithmic-trading, backtesting, synthesis (Search-dominates-compute + the optimization adjacency), index. No new pages in optimization-algorithms-wiki (cross-linked existing ones). Ran avoid-ai-writing (fixed “robust”/“landscape”). +1 page (→11). Verify GREEN.

[2026-06-30] ingest | TradeMaster (NTU) — comprehensive RL quant-trading platform

Routed from Telegram. New source trademaster (SoftwareSourceCode, T3, github.com/TradeMaster-NTU; Apache-2.0, ~2.9k★) — NTU AMI group’s holistic RL-for-quant-trading platform: 6 modules (multi-modal data / preprocessing / market simulators / 13+ algorithm zoo / 6-axis-17-measure eval / web UI), 5 tasks (portfolio mgmt, algo trading, order execution, HFT, market making), 8 datasets. Sibling/superset of tensortrade — it explicitly contrasts itself as the holistic ecosystem vs TT’s lightweight library (recorded on both pages as the library↔platform spectrum). Integrated: deepened reinforcement-learning-trading (two scales of implementation); advanced TWO synthesis open-Qs — “do they compose?” (TradeMaster spans data→sim→policy→eval in one ecosystem = partial yes) and “where’s the evidence?” (systematic eval toolkit + published-paper backing = rigor moved from “none” to “first-party, not neutral”). FinAgent/Market- GAN bridge to financial-time-series-foundation-model/kronos-financial-foundation-model. New entity nanyang-technological-university (CollegeOrUniversity, maintainer). Ran avoid-ai-writing. +2 pages (→13).

[2026-07-15] ingest | Smolinski’s H1-2026 after-action review (Business Insider, hub-routed, Telegram)

Ingested a BI profile (Kathleen Elkins) of discretionary options trader Erik Smolinski‘s 19-slide H1-2026 after-action-review deck. Content: he runs trading as a business (quarterly earnings report = AAR); regime analysis (the deck’s centerpiece) splits H1+early-July into 5 regimes — low-vol grind / ~9% geopolitical drawdown / V-recovery / June mega-cap pullback / dispersed rebound — with speed of transition the real signal; broadening leadership (Mag-7 regime out of vogue, small caps leading); dip-buy speed (only ≥3% pullback fell 9.1%, recovered in 11 trading days, 3 faster than the 2021–25 median); H2 risk = a renewed rate rise (futures priced a possible Sept Fed hike) flipping breadth back to mega-cap/tech; a methodology self-critique slide (data = yfinance / FRED / Cboe SKEW); and a retail prescription — quarterly checkup + benchmark panel SPY/QQQ/IWM/TLT/GLD. Self-reported perf +33.44% vs 10.79% S&P (H1), +57.01% vs 21.49% (TTM) — unaudited, flagged. Routing: boundary case — a discretionary human trader, not a system. Routed here (most-specific financial spoke) on the evaluation clause of the domain; it’s the spoke’s first source where the eval layer is a human narrative review, the concrete first step of the “broaden toward general quantitative finance” growth edge. Runner-up spoke: none (no other financial spoke). New pages: source smolinski-h1-2026-aar (Article, T2 — reputable publisher, self-reported numbers); concepts after-action-review + market-regime-analysis; Person erik-smolinski. Entity discovery: subject Smolinski paged; publisher (Business Insider) + author (Elkins) deferred (thin, borderline single source). Synthesis: added a human-discretionary evaluation bullet to “shape of the field” and reinforced the self-reported-performance tension. avoid-ai-writing self-pass (clean). +4 pages, 3 spine files updated.

[2026-07-26] ingest | Multi-Threaded Trading Robot with Machine Learning (Koshtenko, MQL5, 2026-07-23)

Routed here by the hub (runner-up: optimization-algorithms-wiki, declined by its own rule — that spoke takes the optimizer as subject and excludes trading strategies; here the trading system is the subject). T3 author-published, same tier as the founding bots. New: multithreaded-ml-trading-robot (source), supervised-learning-trading and synthetic-market-data (concepts), yevgeniy-koshtenko (entity). Updated: backtesting (a section on distrusting your own simulator), market-regime-analysis (the automated GMM-clustering counterpart to Smolinski’s hand-drawn regimes), synthesis (decide-layer split now three ways; a new data-scarcity bullet; the evidence open question sharpened). Three things this adds that the corpus lacked:

A supervised decide layer. Kronos forecasts, TensorTrade learns a policy, banbot runs a coded one — nobody was fitting a classifier to manufactured labels, which is the most common practitioner shape. Its hard part hides as preprocessing: markets have no labels, so his label is “would a 300/800 bracket have won?” and changing the bracket changes the game the model learns.
Data scarcity as the constraint under the ML layer, with augmentation (1,000 hours → 5,000 examples, including inversion) and proposed GANs. Pairs with trademaster‘s Market-GAN — two independent reaches for synthetic history, and the circularity problem noted on the new page.
An author who discounts his own results. 64–65% accuracy, encouraging curve, then “I do not really trust homemade Python testers” and no live validation. Every performance number in this corpus comes from a simulator its own author wrote; he’s the first to say so. Also recorded: the concurrency trick is not MQL5 concurrency at all — Python threads per symbol with MT5 demoted to an order endpoint, because the platform’s execution model is single-threaded. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-27] ingest | awesome-systematic-trading (paperswithbacktest, GitHub)

Routed here by the hub (runner-up: none — a systematic-trading catalog is this spoke’s subject end to end; the optimization/pricing library categories are cross-spoke context, not a second home). T3 community link catalog, ~8.8k★, no license file, last pushed 2025-01-22 — 18 months stale, recorded on the page. New: awesome-systematic-trading (source), vectorized-backtesting (concept), paperswithbacktest (entity). Updated: backtesting (two engine families), event-driven-trading (the vector-based contrast), algorithmic-trading (the supporting layers the predict/decide/execute stack omits), synthesis (new bullet + two open questions moved). What it adds that no prior source did:

The map instead of a system. Every earlier ingest was one tool or one practitioner. This names the field’s incumbents — zipline, backtrader, QuantConnect Lean, vnpy, nautilus_trader, vectorbt, Freqtrade — none of which had a page here, and its taxonomy is itself a claim about how the field divides.
Vector-based vs event-driven as the top-level backtester split. The spoke had event-driven as the bias-safe architecture with nothing on the other side; the vectorized family trades the lookahead guarantee and live parity for sweep throughput, which turns the corpus’s “search dominates compute” line into an architecture choice rather than a speed benchmark.
A strategy table that keeps its losers — paper + Sharpe + volatility + rebalancing + runnable QuantConnect code, including Short Term Reversal with Futures at -0.05. Still not the independent referee the evidence question wants (the Sharpes are the papers’ own, and the owner sells the strategy database), but the first entry here that reports a negative result. Also recorded: strategies have migrated off the repo to paperswithbacktest.com — free catalog as funnel, paid database as product. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-08-01] ingest | Quant Bible (MIT Sloan Business Club, 2022, PDF — hub-routed, Telegram)

Source: raw/quant-bible-mit-sloan.pdf (51 pp., pdfTeX, dated 2022-01-17), arrived as a Google Drive link. Held in raw/ — the first document this spoke holds rather than links; Drive URLs rot. Tier T3: student-written, self-published, no review, and its distinctive content (per-firm question banks) is recalled interview material. Runner-up spoke: ../machine-learning-wiki. Created: quant-bible-mit-sloan (source), market-making, quant-interview, quant-trading-firm (concepts), mit-sloan-business-club, massachusetts-institute-of-technology (entities). Updated: algorithmic-trading (market making as the odd-one-out strategy + a “who runs it” section), index, synthesis (two new bullets, microstructure gap narrowed, two cross-spoke adjacencies added). What it adds that no prior source did:

The people and the institutions. Every earlier ingest is a codebase or a lone practitioner. This is the first source about who trades — prop shops vs. quant funds vs. bank desks, and the QT / QR / SWE role split. It arrives refracted through recruiting, because that’s the only thing these firms publish.
Market making as an obligation. The spoke had it as one word in a strategy taxonomy. The guide defines the job (always be available to buy or sell at your own quotes), the pay (the spread), the control (skew against your inventory), and the adversary (adverse selection) — the corpus’s first real microstructure content.
The hiring idiom is statistics, not ML. Regression, econometrics, causal inference, stochastic processes; ESL as canon; nothing asking for a neural network. Measurable distance from what the spoke’s software sources actually build, part vintage (2022) and part whiteboard constraint. Deferred deliberately: the named student contributors (private individuals, byline-only evidence) and the ~27 firms as individual Organization nodes (the source supports a landscape, not 27 pages). Verify deferred per hub policy (content-only, no page moves). avoid-ai-writing run.

[2026-08-03] ingest | Algorithmic Trading: Winning Strategies and Their Rationale (Chan, Wiley 2013)

Routed from the hub (Telegram, a Thalia bookseller PDF). New: chan-algorithmic-trading (T2 source summary), mean-reversion (DefinedTerm/strategy-family), ernest-chan (Person), john-wiley-and-sons (Organization). Updated: backtesting (a new section on failures that survive correct code), algorithmic-trading (two-camps taxonomy beside the seven-families one), strategy-optimization (search less vs guard the search), market-regime-analysis (regime shift as the pitfall implementation care can’t remove), synthesis (first internal contradiction

two open-question updates + a new open question), index. Fetch note: the PDF 403’d through Cloudflare on a plain curl (HEAD returned 200/application/pdf, GET returned the challenge page). Refetched with browser headers + a thalia.de referer; 7.4 MB, 15 pages, extracted with pypdf. Scope is front matter + preface only — no chapter text — and that limit is stated in the frontmatter, on the page and in the tier justification. Substance: eight chapters, four on mean reversion, two on momentum, one each on backtesting/ execution and risk. Mean-reversion apparatus named in full (ADF, Hurst, variance ratio, half-life; CADF and Johansen for cointegration; linear/Bollinger/Kalman for trading it) across stocks, ETF pairs and triplets, currency pairs, futures calendar and intermarket spreads, and the VX future. Two findings. (1) First internal contradiction in the spoke. Chan’s organizing principle is “simple and linear strategies, as an antidote to the overfitting and data-snooping biases that often plague complex strategies”, with explicit scorn for the data-mining approach. Every other source here bets on more model — Kronos, TensorTrade, TradeMaster, supervised-learning classifiers, Koshtenko’s ensemble — and strategy-optimization exists to search hard. Recorded as a live methodological fork with the date asymmetry noted (2013, pre-deep-learning-wave) and neither side adjudicated. (2) The failure catalogue the evidence question wanted. Data-snooping, survivorship, primary vs consolidated quotes, venue-dependent FX quotes, short-sale constraints, futures continuous-contract construction, closing vs settlement prices, regime shift. Six of seven are properties of the data, so the spoke’s engine-side sources address none of them — that’s now on backtesting as a separate section from lookahead bias. Also recorded: the Wiley Trading series blurb (“books by traders who have survived… and have prospered”) is a survivorship filter stated as editorial policy, which is the same selection effect the spoke keeps finding in strategy catalogs, one level up. T2 justified: primary text from a trade publisher, but front matter only — nothing here establishes that any strategy works, and the software survey (MATLAB, Deltix, TradeLink) is a 2013 snapshot. Momentum stays a name: two chapters, and the preface cuts off mid-sentence on its theme. Flagged as the obvious next gap in that corner. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-08-03] ingest | Superior Skills (Superior-Trade) — agent skills for live trading

Routed from the hub (Telegram). T3 — vendor-published, first-party backtests on the vendor’s own engine, strategies selected by the party selling access. No independent replication, no out-of-sample window, no Sharpe, no volatility, no trade log. Skills are MIT and inspectable; the backtester, data and execution sit behind an API key, so nobody outside can check the numbers. Not T4 because the risk disclosure is specific and names its own losing conditions. New pages: superior-skills (SoftwareSourceCode, source), agent-deployed-trading (DefinedTerm/practice). Updated: backtesting (the specimen), market-regime-analysis (a regime gate shipped as a component, and regime-complementarity as portfolio construction), synthesis, index. Dedup: mean-reversion, backtesting, market-regime-analysis, algorithmic-trading, strategy-optimization all already paged — linked and updated, none duplicated. Gap-relevance: lands on the standing “Where’s the evidence?” question with the sharpest specimen yet, and opens a new axis (agent as deployment surface) the spoke had no source for. The finding, stated plainly. The repo advertises a “validated” strategy — Donchian Strong-Regime, BTC, 162 days — at 6 trades, 100% win rate, 0% maximum drawdown, +6.69%. Six trades is not a sample; a perfect record over six is roughly a 1-in-64 coin flip before any selection over strategies and parameters; and a 0% drawdown is what a rule that almost never trades looks like. Recorded on backtesting as a reading rule for the whole corpus: check the trade count before the return. Fair to the source: its risk disclosure is better than most here — it names the single 162-day window, declines to annualize, and says ungated directional strategies are “demonstrably fragile.” The failure is not concealment but attaching “validated” to a result a walk-forward split can’t even be run on. The second strategy (84 trades, 4 pairs, 65.5% win, 18.5% DD) is a claim with enough events to argue about, and is recorded as the contrast. Synthesis: new section “The agent became the deployment surface.” Two claims — the last human check becomes an approval prompt (where a trade has no diff and exit-all doesn’t unwind closed positions), and open skills as a customer-acquisition channel: MIT interface, paid backtester/data/execution, with the same party writing the strategy, running the validating backtest, and earning on deployment. Recorded against the spoke’s new complexity-vs-simplicity fork: the Donchian rule is exactly the simple, linear kind chan-algorithmic-trading argues for, so this is not a point for the ML side. The correction it forces is that a simple model is no defence against a sample of six — Chan’s antidote addresses the model class, and nothing in the corpus addresses the evidence. Cross-spoke: the Agent-Skills packaging is agentic-tooling-wiki’s subject, and the guardrails material on what an approval prompt is worth lives there. Referenced in prose rather than as [[wikilinks]] — this spoke has no cross-wiki link precedent and the link checker is spoke-local. Entities: Superior Trade (the vendor) not paged — evidence here is one repo and a marketing site. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-08-03] ingest | Pseudo-Mathematics and Financial Charlatanism (via research pass)

Hunted, not routed — the hub Research Pass took this spoke’s top ## Most wanted edge (“first T1: auto-edge, zero-T1 spoke — needs a peer-reviewed backtesting-bias paper”). WebFetch was hard-blocked (403) by ams.org; the firecrawl fallback read the PDF, so the ingest is from the paper itself rather than a summary of it. T1 — peer-reviewed, Notices of the American Mathematical Society 61(5), 2014. The spoke’s first T1 source and the first written by someone with nothing to sell. New: pseudo-mathematics-financial-charlatanism (source), backtest-overfitting (concept). Updated: backtesting (a hazard no correct engine can fix), strategy-optimization (the loop that manufactures N), synthesis (the evidence question, answered rather than moved), index. Entities: 6 created — david-h-bailey, jonathan-borwein, marcos-lopez-de-prado, qiji-jim-zhu, lawrence-berkeley-national-laboratory, american-mathematical-society. Queried the entity index first; all eight candidates returned NO MATCH. Western Michigan University and Guggenheim Partners left as prose under the relevance gate — one affiliation mention each, no second connection. Gap-relevance: closes the spoke’s oldest open question (“Where’s the evidence?”) by reframing it. The question assumed a missing referee; the paper shows the missing item is the trial count N, without which a reported Sharpe carries no information at all. First ## Growth edges section added to synthesis per the new hub convention, ranked, top edge = a source that reports its N. avoid-ai-writing run over the new prose.

[2026-08-05] ingest | Fincept Terminal

Routed from the hub (runner-up: agentic-tooling-wiki). fincept-terminal — an open-source Bloomberg-class finance terminal (C++/Qt, embedded Python; ~29,613★), T3, custom NOASSERTION licence. Created fincept-corporation (Organization; entity-index no match ≥ threshold). Budget: 1 of 8.

A broad multi-spoke source, routed whole, dominant substance ingested, facets noted (per HUB). The terminal spans predict → decide → execute plus data and research: QuantLib suite, factor-discovery/ ML/HFT/RL “Quant Lab”, DCF/portfolio/VaR, real-time + algo + paper trading, 16 broker integrations. That trading/quant stack is the dominant in-scope chunk and the reason it lands here. Cross-spoke facets recorded on the page, not fragmented: the 37 investor-persona AI agents + MCP node workflows → agentic-tooling-wiki (logged runner-up); multi-provider LLM support → llm-providers- wiki; macro/market-data connectors (FRED/IMF/World Bank) → a terminal facet no spoke owns.

Synthesis gained a section: it is the first whole-terminal source (all layers at once), and it extends rather than eases the spoke’s evidence problem — it ships every overfitting tool (factor mining, ML, HFT, RL) and reports no validation discipline (no MinBTL, no held-out protocol, no trial count N). The superior-skills “6 trades, 100% win” failure shape on a much larger surface. Also marks the agent-as-terminal turn for agent-deployed-trading: persona agents inside a full application, the trade still with no diff to inspect.

[2026-08-06] ingest | staskh/trading_skills — the second trading skillpack, and a correction

Routed from the hub (Telegram, curator). Runner-up spoke was ../agentic-tooling-wiki, which owns the skills standard and the packaging; the substance here is options trading, and this spoke already held superior-skills and agent-deployed-trading, so it came here and the packaging angle is a cross-link.

New page staskh-trading-skills (T3, repo README). 25 SKILL.md skills, MIT, 324★: market data, indicators, Greeks, PMCC and whale scanners, Interactive Brokers portfolio reads, PDF reports. Yahoo Finance data, ~15 minutes delayed.

It earns its place by being a control case rather than a second example. agent-deployed-trading had concluded from one artifact that an open trading skillpack is a customer-acquisition channel for a paid platform. This one is open, has no platform, and nobody earns when a trade fires — so that paragraph got a correction: openness says nothing about incentives, what the pack needs a key for does.

The better finding is that the two fail on opposite sides of the spoke’s evidence problem. Superior ships strategies and calls one “validated” on six trades. staskh ships no strategies and no backtester, so it makes no claim at all — which is not an improvement, because nothing then sits between a scanner hit and a position except the trader, and the agent’s fluency does the persuading. Folded into agent-deployed-trading and synthesis.

Also recorded: IB is read-only by default with an opt-in switch to read-write. A genuine default-safe boundary, and still a single decision made once that covers every later order. And a caveat the README treats as a footnote — Greeks and unusual-activity scans computed off quotes that are a quarter-hour stale.

No entity node for staskh. The handle is all the source gives; ../ENTITIES.md is evidence-only and a Person page would have been a name and nothing else. Deliberate omission, not an oversight.

[2026-08-07] ingest | 5 Best AI Trading Bot Platforms in 2026 (Innovation & Tech Today)

Routed from the hub (Telegram). T4 and ingested anyway — the soft gate never refuses a curated source; the weakness goes on the record instead.

New pages: ai-trading-bot-platforms-2026 (source summary), trading-bot-platform (concept), innovation-and-tech-today (publisher entity).

What the source is. A ranked roundup of five retail bot platforms. Observable: byline “Marketing/IT Dept.”; every link in the article points at the #1 pick with a ?sf=x777 tracking parameter, including the phrase “best AI trading bot platforms” before any platform is named; that platform also wins the comparison section, the conclusion and three of five FAQ answers; the only commercial specifics in the piece ($99 starter credit, 10% discount) are its; no disclosure anywhere; nothing tested. Paid placement is the inference and is labelled as one.

Why it earns pages anyway. It is the spoke’s first look at the retail market — the corpus is otherwise open-source frameworks, textbooks and papers — and its control axis (managed → configurable → exchange-native → signal-only) is a sound cut that maps onto the predict/decide/execute stack. Folded into synthesis with the finding that matters: the article never mentions Sharpe, out-of-sample testing, slippage or capacity. Against pseudo-mathematics-financial-charlatanism‘s seven-configurations result, the gap between how these systems are sold and how they would be evaluated is the useful content.

Entities: 1 created (innovation-and-tech-today). SaintQuant, 3Commas, Pionex, Cryptohopper and Trade Ideas were deliberately not paged: everything known about them arrived through an advertisement, and a vendor node built from ad copy would launder it into the graph. They are named on trading-bot-platform as unverified mentions. Revisit if an independent source covers any of them.

[2026-08-08] ingest | The Deflated Sharpe Ratio (Bailey & López de Prado, JPM 2014)

Research pass (20-edge run), hub edge #15. deflated-sharpe-ratio new, T1.

The wiki already named this paper twice without holding it — marcos-lopez-de-prado calls it “the companion correction” and backtest-overfitting describes the disease. Now it is a source.

What it adds beyond the concept: the trial count is an input to the statistic. A Sharpe ratio selected out of many configurations is the maximum of a sample, biased upward by construction; the DSR deflates it by the expected maximum under the null, correcting selection bias and non-normal returns, with sample length entering as well.

Edge #1 marked half closed. The correction is here; a source that states its own N is not. The sharper version of the gap, now written into the edge: a backtest without N is not under-documented, it is uninterpretable, because nobody can deflate it.

Not read this cycle, noted for the next: Spurious Predictability in Financial Machine Learning (arXiv 2604.15531) looked on-topic for the same edge.

2026-08-11 — ingest: Black & Scholes (1973) + option pricing (quality cycle)

Coverage edge 4 half closed. The spoke had no page on options at all, which left it unable to describe most of what professional desks trade.

black-scholes (T1, Journal of Political Economy 81(3):637–654, 1973) — the spoke’s third T1 and its oldest source by four decades. Held: the seven “ideal conditions,” the hedged position of one share against 1/w₁ options, the no-arbitrage step to the differential equation, the closed-form call formula, and the result that the expected return on the stock does not enter it. Also x·w₁/w > 1 — an option is always more volatile than its stock.

option-pricing (concept) — the layer the predict→decide→execute stack lacks. Five inputs, four observable; volatility inverted out of the market price is what an options market really quotes. Delta enters as Black and Scholes’ hedge ratio rather than as a risk statistic. Gamma, vega, theta and rho are recorded as a gap, not written up — the corpus has no source defining them, and staskh-trading-skills only lists them as a feature computed off ~15-minute-delayed Yahoo data.

Extraction note. The Princeton mirror of this paper (cs.princeton.edu/.../black_scholes73.pdf) is a JBIG2 scan with no text layer — pypdf returns the JSTOR cover page and nineteen empty pages. The SFU copy (sfu.ca/~kkasa/BlackScholes_73.pdf) is the same scan with a text layer and extracts cleanly at ~45k characters. A failed extraction says something about the file, not the paper.

The finding, and it is about this corpus rather than about options. Black and Scholes report their own empirical test finding systematic deviations from their formula — buyers overpay, widest on low-risk stocks — and then state that transaction costs make the deviation untradeable. A negative result about their own model, published. Nothing else in this spoke does that. Written up in synthesis as The corpus’s oldest source is the only one that reports where it fails, with the caveat that a JPE paper and a GitHub README answer to different reviewers.