Authority signal
Corrections Log
Every factual correction, clarification, and retraction made to AI Fin Hub Research content is logged here, dated, and linked from the affected article.
Policy
When a factual error is found in a published article, tool, or benchmark, the content is corrected inline and the "Last updated" timestamp is bumped. Material corrections are flagged at the top of the affected page and logged here. Corrections are never silently edited. History is preserved.
-
Corrected
A factual error was fixed.
-
Clarified
Wording was sharpened to prevent misreading (no factual change).
-
Retracted
The claim or article was removed because it could not be substantiated.
Log
- correctedTiingo vs EODHD vs Financial Modeling Prep 2026
Gave Tiingo Power's unique symbols per month as exactly 110,329 while the page's recorded figure was 110,328. Tiingo's pricing page shows a live count of the securities it covers, which moves daily (110,424 on 2026-11-02), so the article now says about 110,000 and the recorded figure carries the value and date it was read.
- clarifiedPolygon.io vs Databento vs Alpaca Market Data API 2026
Described Massive Stocks Advanced as "real-time SIP", citing Massive's pricing page, which says real-time data and does not mention the SIP. The article now says real-time, as the cited page does.
- correctedLong-Context Financial Question Answering: SEC Filings & RAG
The opening section said EDGAR holds about 18 million filings and takes 1.2 million a year, that a large-cap 10-K is 80 to 140 printed pages and 60,000 to 150,000 tokens, and that Item 1A alone can exceed 20,000 tokens, citing model documentation that states none of it. It now gives the SEC's own figures (about 4,700 filings processed a day, about 40,000 new filers a year) and the 10-K token range from the page's measured EDGAR sample.
- correctedState of LLM pricing for finance, Q3 2026
Mistral Medium 3.5, Large 3 and Small 4 were listed with a 128K context window; Mistral's model pages give 256k for all three. The table, data.csv and data.json are corrected, and the source note adds the model pages because mistral.ai/pricing now shows consumer plans.
- correctedState of LLM pricing for finance, Q2 2026
Gemini 2.5 Pro, Flash and Flash-Lite were listed with a maximum output of 65,535 tokens; Google's model pages give 65,536. Corrected in the table, the text, data.csv and data.json.
- correctedArticle listings, RSS and JSON feeds, search
Listings, feeds and search read a separate copy of each article's title, description and dates, and 28 of those fields no longer matched the article pages. Some kept claims the pages had since corrected: the execution-simulation summary still said market impact scales as the square root of trade size (the page uses the Almgren et al. 2005 model), and the calibration-drift title was cut off mid-phrase. Article pages now read the same copy as the listings, so the two cannot differ.
- correctedDeepSeek V4 for Finance 2026: SEC Filing Extraction Cost
The legacy model names deepseek-chat and deepseek-reasoner were said to retire at 15:59 UTC on 2026-07-24. DeepSeek's change log gives the date only, so the page now says they were discontinued on 2026-07-24.
- correctedTrading System Blueprinter
The data-source option was labelled Polygon.io, which rebranded as Massive on 2025-10-30; it now reads Massive (formerly Polygon.io). Saved links keep working: the option's id is unchanged.
- correctedSEC Filing Chunk Optimizer
The 10-K preset (120,000 tokens) was described as a median measured from EDGAR filings with the cl100k tokenizer, which we cannot reproduce. It now uses the median of the 10-K main documents we measured on EDGAR (about 91,000 tokens at 4 characters per token, sample published at /data/filing-length-sample.json); the MD&A, notes and transcript sizes are now labelled as estimates.
- correctedFinancial QA LLM Benchmarks 2026: FinanceBench & Fin-RATE
The page said a large-cap 10-K resolves to roughly 120,000 tokens without a source. The cost table still prices a 120,000-token read, now described against our published EDGAR sample: above the median 10-K main document (about 91,000 tokens) and within the large-cap range, with the longest, a large bank's, at about 273,000.
- correctedReading Financial Filings With LLMs: 2026 Playbook
The filing sizes (10-K 60-150K tokens, 10-Q 20-60K, 8-K 1-5K) had no source. They are now computed from a sample of recent filings we measured on EDGAR and publish at /data/filing-length-sample.json: 10-K main documents about 23K to 273K tokens (median 91K), 10-Qs about 21K to 189K (median 35K), an 8-K form about 1K before exhibits.
- correctedMarket Data API Cost Calculator
An FAQ still referred to Polygon, which rebranded as Massive on 2025-10-30; it now uses the current name, as the tool does.
- correctedFinance MCP Directory
The FAQ and how-to called the catalog a 2026-04-20 snapshot after it had been re-checked; both now read the catalog's own checked date.
- correctedBroker API Comparator
The FAQ said the broker snapshot was last checked on 2026-04-23; the data was re-verified on 2026-10-19, and the FAQ now reads the date from the dataset so the two cannot diverge again.
- correctedAgent Cost Envelope Calculator
The FAQ said Anthropic bills cache reads at about 0.1x the input rate; for the current Claude Opus 5.5 and Sonnet 5.5 it is 0.05x (0.1x on Haiku 5.5).
- correctedFinancial Document Token Estimator
The document presets were hand estimates far below real filings: the 10-K preset was 72,000 characters (about 18,000 tokens) against a median of 365,600 characters (about 91,000 to 104,000 tokens) in 15 recent 10-K main documents we sampled from EDGAR on 2026-10-19, and the 10-Q preset was a quarter of its measured median. The SEC presets are now computed from that sample, which is published with the script that produced it; the 8-K preset is the form without exhibits. The paste box was capped at 50,000 characters, too small for any 10-K in the sample; it now takes 1.5 million. The page description and how-to counted 10 models and gave context windows of 200K to 1M tokens; the tool shows 13 models with windows of 400K to about 1.05M, and the FAQ and how-to now read both from the model catalog.
- correctedHow to Read Options Greeks
The guide called delta a rough probability that the option finishes in the money without saying how rough. In Black-Scholes that probability is N(d2), a little below a call's delta of N(d1); the guide now gives both.
Merged into Options Greeks Calculator (Black-Scholes) on 2026-10-21.
- correctedHow to Build a RAG Pipeline Over SEC Filings
The guide said smaller chunks drive embedding cost because they mean more chunks and more embeddings. Embedding is billed per token, so splitting the same text into smaller pieces costs about the same; smaller chunks cost more only through overlap, which repeats a larger share of the text as chunks shrink. The cost step and the chunk-size answer now say so.
Merged into SEC Filing Chunk Optimizer on 2026-10-21.
- correctedState of AI Market Data 2026
The LLM cost table priced superseded models (Sonnet 4.6, Opus 4.7, GPT-5, Gemini 2.5) with figures that did not follow from its own stated loop: most rows' monthly costs implied about 3,800 calls, one row 1,226, and two per-call figures matched older price lists. It also said Opus is better on under 10% of prompts, with no source. The table is now computed from the current model catalog for 1,265 calls a month, with the arithmetic stated, and the quality claim is removed.
- correctedAlpha Vantage Alternatives (2026)
FMP's entry said its premium prices are not consistently documented and should be confirmed, while listing them ($29, $69 and $139 a month); the contradiction is removed. The intro dated all prices to May 2026 although several rows were re-checked later; Alpha Vantage and FMP pricing are now cited.
- correctedEODHD Alternatives (2026)
FMP's entry said its premium prices are not consistently documented and should be confirmed, while listing them ($29, $69 and $139 a month); the contradiction is removed. Massive's Stocks Advanced plan carries 20+ years of history, not 15+.
- correctedFinnhub Alternatives (2026)
FMP's entry said its premium prices are not consistently documented and should be confirmed, while listing them ($29, $69 and $139 a month); the contradiction is removed.
- correctedTiingo Alternatives (2026)
FMP's entry said its premium prices are not consistently documented and should be confirmed, while listing them ($29, $69 and $139 a month); the contradiction is removed. Massive's Stocks Advanced plan carries 20+ years of history, not 15+.
- correctedPolygon.io (Massive) Alternatives (2026)
An FAQ priced FMP Premium at $49 a month; it is $69 a month, or $49 billed annually.
- correctedWalk-Forward vs K-Fold Cross-Validation
The page credited combinatorially symmetric cross-validation with purging and embargoing. CSCV (Bailey et al.) estimates the probability of backtest overfitting; purging and embargoing belong to López de Prado's purged k-fold and combinatorial purged cross-validation, now named and cited.
Merged into Walk-Forward Validation Visualizer on 2026-10-21.
- correctedSharpe Ratio vs Sortino Ratio
The page said the two ratios agree for symmetric returns and read a Sortino above Sharpe as mostly-upside volatility. For symmetric returns, downside deviation is at most 1/√2 of the standard deviation, so Sortino is at least about 1.4 times Sharpe; skew shows in departures from that baseline. Downside deviation is now defined as the root mean square of shortfalls, not the standard deviation of below-target returns.
Merged into Sharpe Ratio Calculator on 2026-10-21.
- correctedSimple vs Exponential Moving Average
An FAQ said the usual EMA convention matches an SMA "roughly twice as fast"; the 2/(N+1) smoothing factor gives an N-period EMA the same center of mass as an N-period SMA, so both trail a steady trend by (N-1)/2 periods. The EMA's advantage is now described as reacting sooner to fresh moves, not less lag overall.
Merged into the tools index on 2026-10-21.
- correctedFixed vs Trailing Stops
A citation, "Stop-Loss Rules and Their Performance" at an SSRN address, could not be matched to that title; it is replaced by Kaminski and Lo, "When Do Stop-Loss Rules Stop Losses?" (Journal of Financial Markets, 2014).
Merged into the tools index on 2026-10-21.
- correctedArithmetic vs Logarithmic Returns
The FAJ citation (Jacquier, Kane and Marcus, 2003) linked to the journal's home page; it now links the article's DOI.
Merged into Returns Distribution Analyzer on 2026-10-21.
- correctedRealized vs Implied Volatility
Two citations shared one link: Carr and Wu's "Variance Risk Premiums" pointed at Bollerslev, Tauchen and Zhou's article. Each now links its own DOI.
Merged into Returns Distribution Analyzer on 2026-10-21.
- correctedMonte Carlo vs Historical VaR
The "Filtered Historical Simulation" citation linked to an address that resolves to no such article; it now cites Barone-Adesi, Giannopoulos and Vosper's FHS backtesting paper (European Financial Management, 2002).
Merged into VaR Backtest — Kupiec & Christoffersen on 2026-10-21.
- correctedDollar-Cost Averaging vs Lump Sum
The page said lump-sum wins because markets rise more often than they fall; it wins in expectation whenever the portfolio is expected to beat cash, which a positive mean return does not make the same thing. A citation to an FAJ article, "Nibbling on Multi-Asset Markets", could not be found and pointed at the journal's home page; it is replaced by Cho and Kuvvet (Journal of Financial Planning, 2015).
Merged into the tools index on 2026-10-21.
- correctedFinance MCP Directory
The catalog missed TradeStation's official MCP connection, released January 2026 (OAuth 2.0 with PKCE; orders are sent only after confirmation); it is now listed and graded D. The page twice said "Catalog snapshot of 2026-04-20" although the catalog was re-checked on 2026-10-18.
- correctedState of AI Market Data 2026
The MCP table is now rendered from the Finance MCP Directory's catalog instead of a hand-copied table, which adds TradeStation's official connection. The findings' counts follow: one of seven servers that can trade documents idempotent order submission, and four of seven official servers are hosted with no code to audit.
- correctedAlpaca API Alternatives (2026)
The page said Tradier, Interactive Brokers and TradeStation have no official MCP server; Tradier's hosted server launched in December 2025, IBKR's draft-only connector in July 2026 and TradeStation's connection in January 2026. It also called Alpaca's multi-leg options tooling limited; Alpaca accepts multi-leg orders such as spreads and iron condors when every leg is covered within the same order.
- correctedIBKR API Alternatives (2026)
The page said IBKR, Tradier and TradeStation have no official MCP server (IBKR's connector drafts trade instructions; Tradier's and TradeStation's can place orders), and that Tradier's published rate limits vary by source; Tradier publishes 60 trading requests a minute in production.
- correctedHow to Calibrate Probability Forecasts
The guide read every point below a reliability diagram's diagonal as overconfidence; below the diagonal means forecasts were too high, which is overconfidence only above 50%. It described the Brier decomposition as calibration, refinement and uncertainty; Murphy's decomposition is reliability, resolution and uncertainty. It also said isotonic regression can correct non-monotonic distortions; it fits monotonic ones only.
Merged into Calibration Dojo on 2026-10-21.
- correctedHow to Validate a Trading Strategy
The guide said a raw Sharpe of 1.2 can deflate below 0.5 once a few hundred trials and fat tails are priced in, without the inputs. With three years of daily returns, the site's deflated-Sharpe engine gives about 0.98 for a single trial and 0.21 for the best of 300; fat tails change the result little, so the guide now attributes the drop to the trial count. Kupiec's 1995 paper is now credited to the Journal of Derivatives.
Merged into Walk-Forward Validator on 2026-10-21.
- correctedBid-Ask Spread
The page defined realized spread with an absolute value; Regulation NMS defines it as a signed measure that turns negative when the price moves through the trade. It also advised using realized spread as a liquidity taker's cost, when effective spread is that cost, and said charging the quoted 2 cents overstated the example's 1-cent cost by 50%; it is double. The time-of-day claim now cites McInish & Wood 1992.
Merged into Execution Simulator on 2026-10-21.
- correctedLatency Arbitrage
The example put the Chicago-New York round trip at about 400 microseconds; Budish, Cramton and Shim (2015) report 13 ms by fiber and 8.1 ms by microwave. Unsourced figures (six figures a day, hundreds of millions in infrastructure, tens of millions per venue) were removed; the example now uses the paper's measurements: ES-SPY windows of 97 ms in 2005 and 7 ms in 2011, and a prize of about $75 million a year.
Merged into Execution Simulator on 2026-10-21.
- correctedLook-Ahead Bias
The page suggested shuffling random subsets of the time series as a look-ahead test, which does not isolate look-ahead; it now suggests lagging every signal one more bar. An unsourced claim about Sharpe ratios above 3.0 was removed and the example is labelled illustrative.
Merged into Price-Blind Research Auditor on 2026-10-21.
- correctedMaker-Taker
The example's arithmetic was wrong: the maker-taker gap is $1.00 per 100-share round trip (1 cent per share, not two), and 10,000 round trips a day make $10,000 a day, not $200. The page now states the Regulation NMS access-fee cap ($0.0030 per share, falling to $0.0010 by November 2027 under SEC Release 34-105656), and a citation with an invented title was replaced with Battalio, Corwin & Jennings (2016).
Merged into Execution Simulator on 2026-10-21.
- correctedMCP (Model Context Protocol)
The page listed SSE as an MCP transport; the current specification defines stdio and Streamable HTTP, which replaced the HTTP+SSE transport. An unsourced claim that MCP is becoming the de facto standard was removed.
Merged into Finance MCP Directory + Security Grader on 2026-10-21.
- correctedModel Drift
The page cited a gpt-4-0613 to gpt-4-1106 change as a famous case without a source; it now cites Chen, Zaharia & Zou (2023), who found GPT-4's prime-number accuracy fell from 84% to 51% between the March and June 2023 versions. The monitoring example's figures were presented as measurements and are now labelled illustrative.
Merged into Prompt Regression Tester on 2026-10-21.
- correctedMonte Carlo Simulation
The example reported results for a 'Kelly-sized strategy' without the model, parameters or seed that the page itself says every simulation must report. It is replaced by a simulation run for this page with all of them stated (daily returns of 0.05% ± 1.26%, full-Kelly leverage 3.15x, 10,000 paths, seed 42) and a convergence check at 5,000 paths.
Merged into Drawdown-Recovery Markov Simulator on 2026-10-21.
- correctedOverfitting
The page said 20 trials cut about 0.3 and 1,000 trials about 0.6 from Sharpe expectations, and that live-to-backtest ratios typically land at 0.3-0.5, without a source. The expected best Sharpe under the null is now computed from Bailey & López de Prado's formula for five years of daily data (0.85 for 20 trials, 1.46 for 1,000), and the example's results replace figures that were presented as a run.
Merged into Backtest Overfitting Score on 2026-10-21.
- correctedSlippage
The example's impact formula (k × sqrt(Q/ADV) with k = 10) gives 1 basis point, not the 10 shown; it now uses the standard form with daily volatility. The 13 bps entry cost was described as round-trip, the takeaway about linear and square-root impact had the direction reversed, the latency cost was overstated (about 0.3 bps of noise, not 1 bp), and the FAQ credited the square-root law to 'Almenberg'; it now cites Tóth et al. (2011).
Merged into Execution Simulator on 2026-10-21.
- correctedSurvivorship Bias
The citation named Brown, Goetzmann, Ibbotson & Ross for a paper by Elton, Gruber & Blake (1996). Unsourced bias figures (1-2% a year for small caps, 2-4% for hedge funds) were removed, the example's 2.3-point gap over 20 years produces about 52% more ending wealth rather than doubling it, and Lehman's ticker is LEH, not LEHM.
Merged into the tools index on 2026-10-21.
- correctedValue at Risk (VaR)
The example said a 95% VaR is exceeded about five days a quarter; at 5% of about 63 trading days it is about three. The page called VaR the regulator's risk number without noting that Basel's 2019 market-risk standard moved internal models to expected shortfall. A Federal Reserve working paper cited as 'An Overview of Value at Risk' is a different paper on mutual-fund risk taking; the citation is now Duffie & Pan (1997).
Merged into Returns Distribution Analyzer on 2026-10-21.
- correctedWalk-Forward Optimization
An unsourced rule that an out-of-sample to in-sample ratio below 0.5 is a red flag and a claim about what most quant teams do were removed, and the example is labelled illustrative.
Merged into Walk-Forward Validator on 2026-10-21.
- correctedPrompt Injection
A takeaway called capability separation the only architectural defense that holds up; it now says why it is the defense that does not rely on the model resisting the injection, citing Greshake et al. (2023).
Merged into Prompt Injection Tester on 2026-10-21.
- correctedMax Drawdown vs Ulcer Index
The citation for Magdon-Ismail and Atiya's work on maximum drawdown linked to a page that does not exist; it now cites their 2004 Journal of Applied Probability paper on the maximum drawdown of a Brownian motion, which derives how expected drawdown grows with the length of the sample.
Merged into Risk-Adjusted Returns Calculator on 2026-10-21.
- correctedBeta Formula
The steps divided a covariance of 0.00072 by the market variance to get a beta of 0.45, while the worked example and its result used 0.00108 and 0.675. The steps now use the worked example's numbers.
Merged into Risk-Adjusted Returns Calculator on 2026-10-21.
- correctedCalmar Ratio Formula
The steps worked a different example (100 to 144 over two years, a 25% drawdown, Calmar 0.80) from the calculation and result (166.4 over three years, 22%, 0.84); they now follow the worked example. A claim that Calmar above 1.0 is strong for trend-following programmes had no source and was removed, and a ProQuest link to Young's 1991 article returned an error and was removed.
Merged into Risk-Adjusted Returns Calculator on 2026-10-21.
- correctedSortino Ratio Formula
The steps used returns of +3%, -2%, +1% and -4% (a Sortino of -0.224) and then annualized an unrelated 0.30, while the calculation used +4%, -3%, +5% and -2% for 1.92. The steps now follow the calculation, which the calculator reproduces.
Merged into Sharpe Ratio Calculator on 2026-10-21.
- correctedProbabilistic Sharpe Ratio Formula
The steps used 250 observations (PSR 0.942) and the calculation 252 (0.943); both now use 252, which the calculator reproduces.
Merged into Deflated Sharpe Ratio Calculator on 2026-10-21.
- correctedFractional Kelly Formula
The summary said half-Kelly cuts the variance of growth in half; variance scales with the square of the multiplier, so half-Kelly halves the volatility and cuts the variance to a quarter, as the page's own calculation states.
Merged into Kelly Criterion Calculator on 2026-10-21.
- correctedProfit Factor Formula
The example, a 30% win rate with winners four times the size of losers, was called the inverse of a typical trend system and labelled mean reversion; large, infrequent wins are the typical trend-following profile. Ratings of 1.5 to 2.0 as solid and above 2.0 as strong had no source and were removed.
Merged into the tools index on 2026-10-21.
- correctedSharpe Ratio Formula
The standard-error step presented Lo's per-period formula as the standard error of the annualized Sharpe ratio; it is per period and is multiplied by the square root of periods per year to annualize. The step now works the example: about 0.45 annualized, a 95% interval of roughly -0.06 to 1.71.
Merged into Sharpe Ratio Calculator on 2026-10-21.
- correctedGain-to-Pain Ratio Formula
Thresholds of 1.0 as good and 1.5 as very good were attributed to Schwager without a checkable source, and the ratio was called far more robust to outliers than the Sharpe ratio; both were removed.
Merged into Risk-Adjusted Returns Calculator on 2026-10-21.
- correctedTracking Error Formula
The result called 1.66% typical of enhanced-index strategies and said passive funds usually report under 0.5%, without a source; both statements were removed.
Merged into Risk-Adjusted Returns Calculator on 2026-10-21.
- correctedOmega Ratio Formula
The page linked to the returns distribution analyzer, which does not compute an Omega ratio; it now links to the Sharpe vs Sortino calculator, which returns 1.143 for the worked example.
Merged into Risk-Adjusted Returns Calculator on 2026-10-21.
- correctedPrompt Cache Break-Even Formula
Cached reads were described as often 0.1x the input price; that holds for older Claude models, while the newest are billed at 0.05x or 0.025x. The documentation links now point to the current Anthropic and OpenAI pages.
Merged into Token-Cost Optimizer on 2026-10-21.
- correctedCost Per 1K Tokens Formula
Cached input was described as often 10-25% of the input price; Claude cache reads now run 2.5-10% of it depending on the model.
Merged into Token-Cost Optimizer on 2026-10-21.
- correctedTool and article summaries (site-wide)
A meta-description trimmer run on May 10, 2026 also cut the summary text of 19 tools and 3 articles at 155 characters, leaving lines such as "Catches regime." and "cause LLMs." on the tools index, the homepage and the agent tool index. The full summaries are restored.
- correctedKalshi vs Polymarket Arbitrage Calculator
The tool was described as a daily-refreshed scan of arbitrage candidates. It never fetched quotes: its leaderboard ran on three synthetic pairs, and fees were a flat assumption in basis points. It is now a calculator: you enter the asks, and it prices both trade directions with Kalshi's published fee formula (fee schedule effective July 7, 2026, reproduced row for row in our tests) and Polymarket's category fee rates, with net profit and annualized return. The FAQ also said a resolution mismatch makes one leg win and the other lose; with YES on one venue and NO on the other, a mismatch makes both legs lose, which costs the whole stake.
- correctedBroker API Comparator
The comparator listed futures, forex and bonds for Schwab's Trader API. The API places stock, ETF and option orders; futures and forex trade at Schwab on thinkorswim, and a futures order sent to the API is refused. Filters for futures or forex no longer return Schwab.
- correctedBroker API Rate Limits 2026: Interactive Brokers, Tradier
The Schwab section listed futures, forex and mutual funds as asset classes the Trader API covers. Its orders cover US stocks, ETFs and options; futures and forex trade on thinkorswim, not through the API.
- correctedTradeStation API Pricing 2026
The decision list named Schwab alongside TradeStation and IBKR for options plus futures through one API. Schwab's Trader API does not take futures orders; the list now names TradeStation and IBKR.
- retractedFTC vs FCA vs MiCA+DORA Regulatory Cost
Retired the regulatory cost comparator. Its setup hours, monthly hours, registration fees and legal rates were illustrative planning assumptions, not survey data or measured costs, so its totals and its ranking of the three regimes measured nothing. The URL now leads to the BaFin + EU guide, which covers the rules with sources.
- correctedNumeric Precision in LLM Filing Extraction
The synthetic restatement note cited ASU 2024-03 as a revenue-recognition standard; ASU 2024-03 concerns the disaggregation of income-statement expenses. The fictional note now refers to a change in revenue-recognition policy without naming a real standard, and the code example uses a current model id.
Merged into LLM Finance Error Taxonomy on 2026-10-21.
- correctedBond Yield Curve Parsing with Claude Haiku 4.5
Retracted the worked example: the "May 7, 2026 10-year auction" announcement it parsed, with its CUSIP, coupon, yield and bid-to-cover, did not match Treasury's record, and the model output labelled verified was not a recorded run. The May 2026 10-year note was auctioned on May 12 (CUSIP 91282CQQ7, 4.375% coupon, 4.468% high yield, 2.40 bid-to-cover), and the page now uses that record from Treasury's Fiscal Data API, labels the expected output as such, and leads with the API because the data is published in structured form. Also corrected the per-call cost (3,000 input and 400 output tokens at $1 and $5 per million is $0.005, not $0.0008), an Opus cost ratio, an auction cadence of about ten a day (the week of May 11 had ten in total), the framing of FOMC statements as a yield-curve source, and references to a May 7 FOMC statement and a Haiku model card date we could not support.
Merged into Best Economic Data APIs 2026: FRED vs Trading Economics on 2026-10-21.
- correctedExecution Simulator
The simulator's impact coefficients were round illustrative values (0.1 on a square-root term, 0.05 on a linear term), as the page disclosed, and they put typical impact well below published estimates. It now uses the model and coefficients Almgren, Thum, Hauptmann and Li (2005) fitted on US large-cap orders, adds inverse turnover (shares outstanding over daily volume) as an input, and reproduces the paper's Table 3. For the default order (1% of daily volume, 2% volatility, 10% participation) the estimated cost rises from 6.1 to 12.4 bps; faster trading now costs more, as it should.
- correctedExecution Simulation: Slippage and Market Impact
The calibration section attributed the simulator's illustrative coefficients to Almgren et al. (2005), and its own code, at the coefficient it recommended, gave about 74 bps for the worked example while the text said 30. It also described permanent impact as following the square-root law (the 2005 study found it linear in size, with temporary impact following a 3/5 power), called square-root impact the unique arbitrage-free form, and described the Execution Simulator as replaying order books, which it does not. The page now computes every figure with the simulator's engine on the published model (the worked example is about 40 bps), cites the paper's coefficients and Table 3, and drops typical-cost ranges and error bands it had no source for.
Merged into Execution Simulator on 2026-10-21.
- correctedAlmgren-Chriss Optimal Execution: A Worked Example
The worked cost of 19.40 bps came from the Execution Simulator's former illustrative coefficients. With the coefficients Almgren et al. (2005) estimated, the same sell costs about 40 bps; the prose now takes every figure from the engine at build. Halving participation cuts temporary impact by about a third under the fitted 3/5 power, not by half as the page said, and an uncited citation-count claim was removed.
- clarifiedBacktest Overfitting in LLM Trading Strategies: PBO Explained
Removed two effects the page had not measured (that a pre-registered spec lowers PBO by 10 to 20 percentage points, and that the three mitigations together typically halve it) and a claim that passing the three gates guarantees freedom from the failure modes they check. Added the paper's own rejection rule, PBO above 0.05, next to the page's looser screening bands, and the pool composition behind the 30-to-300-candidate result. We re-ran the worked cases with the page's code and seeds; every reported mean and range reproduced.
Merged into Backtest Overfitting Score on 2026-10-21.
- correctedDeflated Sharpe Ratio: Bailey-Lopez de Prado Worked Example
The displayed benchmark formula omitted the factor that scales the expected-maximum term into Sharpe units (the square root of the cross-trial Sharpe variance), and the formula's SR was defined as annualised although the formula needs the per-period figure. The worked tables and the code already used the scaled, per-period form; the formula block and definitions now match them. We re-ran the page's Monte Carlo (mean maximum Sharpe 1.535 against the closed form's 1.47) and recomputed every table row. Also removed a Christie (1982) attribution for the non-normal Sharpe variance that we could not support, skew and kurtosis figures ascribed to XIV without a source, and a claim about where most published strategies' DSR falls; the count of factors tested in the literature is now attributed to Harvey, Liu and Zhu (2016).
- correctedThe BaFin + EU Guide for Retail AI Traders (2026)
Rewritten against primary sources after four errors. The page said publishing buy or sell calls without a licence is unauthorised investment advice; BaFin's guidance says recommendations made to the public are generally not investment advice, and the obligations that do apply are the Market Abuse Regulation's disclosure rules, with undisclosed conflicts potentially amounting to market manipulation. It said the 20,000-euro cap on offsetting derivative losses was still valid; the Jahressteuergesetz 2024 abolished it for all open cases. It cited the Telemediengesetz for the Impressum, which the Digitale-Dienste-Gesetz replaced in May 2024. It cited a BaFin 2025 update and an ESMA 2026 finfluencer note we could not find; the page now cites ESMA's 2021 statement and 2024 warning, BaFin's 2025 guidance note, its January 2026 finfluencer factsheet and its August 2026 article. An unsourced enforcement track record was replaced by the measures the law provides.
- correctedCompliance Audit Trails for LLM-Driven Trade Decisions
Retracted an account of enforcement that the sources do not support. The page said 2024 SEC actions against three advisers turned on missing prompt and model records, with fines from $400,000 to above $1.5 million; the SEC's March 2024 "AI washing" cases were against two advisers, Delphia and Global Predictions, for false claims about using AI, with $400,000 in combined penalties. Claims about a FINRA priorities-letter item and a New York DFS reproducibility requirement are removed; FINRA Regulatory Notice 24-09 is cited instead. The references cited a 2018 SEC no-action letter on S3 Object Lock that we could not find; the support for Object Lock is a Cohasset Associates assessment for AWS, now cited, and the page now notes that Rule 17a-4(f) has accepted an audit-trail alternative to write-once storage since 2022. Also removed an unsourced claim about OpenAI auto-routing incidents and an examination-frequency ranking we had no data for.
- correctedReading Financial Filings with LLMs in 2026
The page said prompt-caching and batch discounts do not stack; Anthropic's batch documentation says they do, with cache hits inside a batch on a best-effort basis. It gave cache reads as 10 percent of the input price, which on the current Claude lineup ranges from 2.5 to 10 percent. The master table's per-document cost bands had been set by hand at June 2026 prices and overstated current costs several-fold (extraction was shown at $0.01 to $0.08 a filing); they are now computed from the model catalog for a stated token profile per task. Removed figures the page had no source for: a three-to-eight-fold chunking saving, 70 to 90 percent fewer classification calls after normalisation, a 95 percent validator catch rate, fine-tuning volume thresholds, and a 2 to 5 percent sampling practice.
- correctedOrder Book Imbalance (glossary)
The entry's only source pointed to the wrong paper: the link resolved to a particle-physics preprint, and the title it gave belongs to a different Cont paper than the authors it named. It now cites Cont, Stoikov and Talreja (2010), Cont, Kukanov and Stoikov (2014) and Stoikov (2018), each linked by DOI. The unsourced claim that OBI is the strongest single short-horizon predictor on most venues is replaced by what Cont, Kukanov and Stoikov report, and the weighted mid-price is no longer labelled the micro-price.
Merged into Order Book Replay Visualizer on 2026-10-21.
- correctedMarket Data APIs Compared 2026
The headline table rated Databento's options coverage as limited; Databento serves the full OPRA feed on metered pricing. The suggested single-operator stack was totalled at $230 to $300 a month, which its own components do not add up to: Alpaca Algo Trader Plus or Massive Stocks Advanced plus Tiingo Power is $129 to $229 before Tiingo's fundamentals add-on. IEX's share of US volume is now given as about 4% (4.43% in July 2026) instead of 2 to 5%, and an unsourced $100 to $500 range for metered Databento spend and an unverified count of Alpaca MCP actions were removed.
- correctedChoosing a Broker API in 2026
The page presented the pattern-day-trader rule and its $25,000 minimum as current; FINRA replaced them with intraday margin standards effective June 4, 2026 (Regulatory Notice 26-10), with a phase-in to October 20, 2027. The FAQ quoted a retired Tradier plan at $0.35 per option contract; Tradier Pro is $10 a month with $0 equity-option commissions. The references cited FINRA Regulatory Notice 21-19, which concerns short-sale reporting, for payment for order flow; the correct notice is 21-23. Removed order-acknowledgment latency ranges and a broker execution-quality ranking that had no measurement or source behind them, and corrected two summary lines that contradicted the body (Tradier covers options as well as stocks; Robinhood has official crypto and agent APIs, but none for coded stock or options strategies).
- correctedSEC EDGAR API 2026
The page said exceeding 10 requests per second earns an HTTP 403 and a roughly 10-minute IP block, citing the SEC. The SEC's pages say only that it may limit further requests from the IP for a brief period, with no duration or status code. The penalty is now quoted as the SEC states it, and the nightly bulk ZIPs are added.
- correctedBroker APIs for AI Agents 2026: MCP Coverage
Corrected the MCP coverage. Tradier's MCP was listed as community-only, though Tradier has run an official hosted server since December 2025; tastytrade was listed as having no official MCP, though it publishes one; and Interactive Brokers and Robinhood have since launched official servers. Grades were recomputed in the Finance MCP Directory. Also corrected that Alpaca options need its paid tier: options trade on every plan, and only real-time OPRA data needs Algo Trader Plus.
- correctedChoosing a Broker API in 2026
Removed the claim that Alpaca ships the only official broker MCP server: Tradier, Interactive Brokers, Robinhood and tastytrade also run official servers, and Alpaca's is the only one that documents idempotent order submission. Corrected the comparison table, which said Interactive Brokers offers no crypto, and updated the Robinhood section for its Agentic Trading MCP. The same MCP correction was made on Schwab API vs Alpaca, Alpaca vs Tradier API, MCP vs Function Calling for Finance, Tradier vs tastytrade API, and the Alpaca and IBKR alternatives pages.
- correctedState of AI Market Data 2026
Corrected the finance MCP table. It listed a NautilusTrader MCP server that does not exist in the NautilusTrader repository, omitted Tradier's official MCP server (live since December 2025), and graded community servers on estimated commit dates and licences. Re-checked against the GitHub API and vendor docs: the table now covers 10 servers, including the official Interactive Brokers, Robinhood and tastytrade servers, with grades recomputed in the Finance MCP Directory.
- correctedRobinhood API: Official or Unofficial 2026
The page said Robinhood offers no official way to trade stocks or options programmatically, which stopped being true soon after publication: Robinhood launched an official Trading MCP for Agentic Trading, which lets an external AI agent place long equity, options and crypto orders in a dedicated Agentic account. The page now covers it, and the Broker API Comparator marks Robinhood, Interactive Brokers and Tradier as having official MCP servers.
- correctedTwelve Data API Pricing 2026
The page said Grow's entry price was $79 a month and called the widely quoted $29 out of date. Twelve Data sells each paid plan at three credit levels, and its pricing page lists Grow from $29 a month (55 credits/min), $49 (144) and $79 (377); Pro from $99 and Ultra from $329. The page now gives the full ladder, and the same correction was made on Alpha Vantage Alternatives.
- correctedBinance vs Coinbase vs Kraken API 2026
Corrected Coinbase Advanced Trade's rate limits, which the page gave as roughly 10 requests/second private and 15 public; Coinbase's documentation gave 30 private (per user) and 10 public (per IP) until at least May 2026, and it now publishes no REST quota. Corrected the claim that $50,000 of monthly volume costs hundreds of dollars on Binance (0.10% of $50,000 is $50). Kraken's entry fees now follow its cross-platform tiers (0.40% maker / 0.80% taker, previously 0.25% / 0.40%), and the Coinbase fee rates are removed because Coinbase shows them only after sign-in.
- clarifiedBinance vs Coinbase API 2026
Coinbase no longer publishes Advanced Trade REST rate limits; its rate-limit page now redirects to the API overview. The page had presented 30 requests/second private and 10 public as current figures. They are now labelled as the last documented values, the page cites the WebSocket limits Coinbase does publish, and Binance's order limits and ban durations are added.
Merged into Binance vs Coinbase vs Kraken API 2026 on 2026-10-18.
- correctedBest News and Sentiment APIs for Trading 2026
Corrected three vendor claims: Finnhub's News Sentiment and Social Sentiment endpoints require a paid plan (its free tier covers company news), Alpha Vantage's Market News & Sentiment endpoint is a premium function rather than part of the free 25 requests/day, and Massive's (formerly Polygon.io) Ticker News does return per-ticker sentiment and reasoning, on every Stocks plan including the free tier.
- correctedDeflated Sharpe in Low-Trial Research Programmes
Corrected every engine figure. The earlier table compared the observed Sharpe with the unscaled expected maximum of N standard normals (2.19 at N = 40) and reported DSR 0.30; the paper scales that maximum by the standard deviation of the trials' Sharpe ratios, and the engine's corrected output is a 1.55 annualized benchmark and DSR 0.633 at N = 40. Also corrected the claim that more data does not help (it does: DSR 0.961 on five years at N = 40), the kurtosis label (the engine takes full kurtosis), the Harvey and Liu citation (42(1), not 41(4)) and a misattributed footnote.
Merged into The Selection-Bias Sharpe Benchmark by Trials and Sample on 2026-10-18.
- correctedFMP API Pricing 2026: Financial Modeling Prep Plans
Replaced the $22 / $59 / $149 prices, which had been taken from third-party reviews, with the figures FMP's pricing page now prints: Starter $29, Premium $69 and Ultimate $139 a month, or $19, $49 and $99 a month billed annually. The same correction was made on FMP vs Alpha Vantage, Market Data APIs Compared, Cheapest Stock Market Data API and in the Data Vendor TCO engine.
- correctedPolygon.io Free Tier Limits 2026
Corrected the free Basic tier from '15-minute delayed' to end-of-day data, and corrected WebSockets from a Developer feature to a Starter feature, per Massive's (formerly Polygon.io) pricing page. The same errors were fixed on Polygon.io Pricing, Polygon.io vs Alpaca and Polygon vs Databento vs Alpaca. Stocks Advanced history is 20+ years, not 15+.
- correctedPolygon.io Options API Pricing 2026 (Massive)
Added the free Options Basic tier that was missing, corrected Options Advanced history from 'tick-level' to 5+ years, and removed an SLA claim that does not appear on the individual plans.
- correctedTiingo vs Polygon vs Finnhub 2026
Corrected Massive's Stocks Business price to $2,499/month ($1,999 is the Options Business price) and removed Finnhub bulk-history dates and a Tiingo IEX-agreement claim that could not be re-verified on the vendors' pages.
- correctedIBKR TWS API Rate Limits 2026
Updated to IBKR's current TWS API documentation: the request limit is market-data lines divided by 2 per second (50 on the default 100 lines), breaches are either rejected with error 100 (session ends after three) or paced depending on a TWS setting, and the 60-per-10-minutes historical rules are documented for bars of 30 seconds or less.
- correctedData Vendor TCO for EU Retail Traders 2026
Corrected prices shown in euros that were US-dollar list prices, removed exchange-fee and compliance-cost estimates that had no source, and corrected the statement that § 83 WpHG record-keeping applies to private retail algo traders (it is addressed to investment-services firms).
Merged into the Data-Vendor TCO calculator on 2026-10-18.
- correctedWalk-Forward Validation Pitfalls in LLM-Generated Strategies
Removed an unsupported '50% chance of a Sharpe above 1.5' attribution, a momentum survivorship figure misattributed to Elton, Gruber and Blake (1996), an arithmetic error in the transaction-cost example and an unsourced bootstrap error multiple; corrected the Almgren-Chriss pages (5-39) and the Elton-Gruber-Blake title.
Merged into Walk-Forward Validator on 2026-10-21.
- correctedExecution Simulator vs Order-Book Replay
Removed Polygon (Massive) Advanced from the list of depth-of-book sources: its self-serve plans stop at NBBO quotes. Corrected Alpaca's SIP feed to top-of-book without depth.
Merged into the Execution Simulator on 2026-10-18.
- correctedTradier API Rate Limits and Pricing 2026
The page said Tradier's commission figures were unpublished and that real-time data was a paid add-on. Both are on Tradier's pricing page and market-data docs; the page now quotes the published commissions and the real-time data rule.
- correctedPolygon.io (Massive) Pricing 2026
The page described the free Basic plan as 15-minute delayed, Advanced history as 15+ years and WebSockets as a Developer feature. Per Massive's pricing page, Basic is end-of-day data, WebSockets and flat files start at the $29 Starter tier, and Stocks Advanced has 20+ years of history; the page now follows it.
- correctedTradier vs tastytrade API 2026
The page said tastytrade publishes no API limits. Its streaming documentation sets caps, and its MCP server applies per-endpoint limits under a 50 requests/second global cap; the page now cites both.
- correctedInteractive Brokers API Pricing and Rate Limits 2026
The page quoted a generic "$1-15/month per exchange" market-data range and a fixed 50 messages per second that always drops the connection. Both are replaced with IBKR's published subscription prices, waiver thresholds and pacing rules.
- correctedSchwab Trader API Status 2026
The page described a flat limit of 120 requests per minute per endpoint. Schwab documents a per-app Order Limit for order-related requests, set at registration, with 120 recommended for most users; the page now describes that throttle.
- correctedPolygon vs Databento vs Alpaca Market Data API 2026
Databento's flat-rate ladder was given as $179 / $1,500 / $4,000 a month. Databento's pricing page lists Standard $199, Plus $1,750 and Unlimited $4,500 (the last two on annual contracts), matching the CME rates effective June 22, 2026.
- correctedFinnhub Free Tier 2026
The page said Finnhub's paid pricing started at about $11.99 a month and was not machine-readable. The pricing page's embedded plan data lists Market Data Basic at $49.99, Standard at $129.99 and Professional at $199.99 a month; the page now quotes those.
- correctedChoosing a Broker API 2026: Rate Limits, Fees, Auth
Corrected the Tradier trading rate limit in the comparison table: the trade scope is capped at 60 requests/minute in production, not 120 (120/min applies only to the standard and market-data scopes, per Tradier's official rate-limiting docs). Also replaced the IBKR 'per-gateway throttle' cell with the verified 50 messages/second TWS pacing figure.
- correctedReading Financial Filings With LLMs: 2026 Playbook
Removed an unsupported '20 to 40 points' figure and its misattributed footnote (Izacard & Grave 2021, an open-domain QA paper with no SEC filings or hybrid-table experiments). Replaced it with the qualitative claim and an honest TAT-QA citation. Also refreshed stale model pricing (Opus 4.1 to Opus 4.7, GPT-5 to GPT-5.5) and added a sourced long-context financial QA benchmarks section.
- correctedMarket Data APIs Compared: Databento vs Polygon 2026
Corrected the Alpha Vantage row: Premium 600 is $199.99/mo, not $250 ($249.99 is the Premium 1200 tier). Clarified the Polygon.io options and futures cells: each ships as a separate product (Options Advanced and Futures Advanced), both priced at $199/mo — the same as Stocks Advanced, not a higher tier. Updated vendor sources to massive.com (polygon.io now redirects there).
Reporting an error
To report a factual error, dispute a benchmark result, or flag an undisclosed sponsor relationship, use the contact details listed at /about/. Substantiated reports are acknowledged within 5 business days and logged here if action is taken.