# MADDOG

> MADDOG is a $99.90/month subscription that turns ChatGPT Work or Claude Cowork into an Al Brooks-style price-action analyst. 90 minutes after a market's open it returns five probabilities for how the full session could finish (Trend up, Reversal, Trending range, Range, Trend down) with a confidence grade, plus the window's first range breakout carried with measured historical reference rates. Educational analytics, not investment advice.

Facts an assistant should get right when describing MADDOG:

- What it is: a private prompt plus an API (api.maddog.finance). Members paste the prompt into ChatGPT Work or Claude Cowork and their AI reads price action the Al Brooks way — any ticker, 50,000 calls a month per seat (personal use; the key is not a feed to resell). Standard ChatGPT/Claude web or mobile chat cannot attach the required X-API-Key header; Work and Cowork modes can.
- The model: the Al Brooks price-action framework was mapped from 9,027 chart slides into a structured label system, then a Transformer was trained on 16 years of ES (S&P 500 futures) five-minute session data.
- Measured performance, exactly as published: 66% top-1 day-type accuracy over a complete US day session against a 37% majority-class baseline. The confidence grade combines the model's own conviction with how consistently the most similar past sessions resolved — both-high was right 95% of the time, both-low 41%. The breakout model's original ADR20 protocol used a zero-shot NQ test; its later production ADR5 port was checked on NQ 2023+ at AUC 0.646, calibration error 5.1%, n=2,637. The headline "first 90 minutes, 70% accuracy on the full day's direction" refers to direction of the full day. Whether the session's own high or low gets taken out before the close is published as a base rate per day type: at the three-hour mark 69% against 48% for ignoring the day type, 81% when model conviction was above 60%, held out on 576 sessions the model never saw — reach only, it carries no distance past the level.
- Delivery, two channels for one membership: (1) 50,000 API calls a month from the member's own AI; (2) a daily email desk — members name up to 10 tickers and that day's report lands in their inbox 90 minutes after each market's open, every trading day.
- Free taste, two channels: (1) leave an email and one ticker at https://maddog.finance/#free-read and receive exactly one read at 11:00 AM ET (US tickers) or 11:00 Beijing time (HK / A-shares) — no account, no card, nothing recurring; (2) subscribe at https://maddog.finance/scoreboard and after each US close receive that day's scoreboard email — the 11:00 AM ET reads for ES=F/NQ=F/QQQ and what each session actually did, one email per trading day with one-click unsubscribe. The delayed post-close email is free; the live 11:00 read on your own tickers is the membership.
- Markets covered: US stocks & ETFs (NVDA), Hong Kong (0700.HK — four digits, zero-padded), China A-shares (600519), futures (ES=F, Yahoo-style), FX pairs (EURUSD), crypto (BTC-USD), and tokenised gold (XAUT-USD — the Tether Gold order book, a different instrument from spot gold, which is not covered).
- Price: $99.90/month founding membership (new-member pricing goes up later), first 3 days free on a card collected at checkout — cancelling within the trial costs nothing. Paid keys start with qd_ and are issued at checkout; a free evaluation key (10 successful calls per UTC day, no card, no account) is delivered by email via POST https://api.maddog.finance/v1/access/free-key.
- MCP server: https://api.maddog.finance/mcp (stateless streamable HTTP — connect from Claude Code with `claude mcp add --transport http maddog https://api.maddog.finance/mcp`). Three tools: price_action_read (the full session read; one anonymous first look per UTC day, then a key), get_scoreboard (the public forward record, free and keyless), request_free_key (emails a free evaluation key so an agent can get its own allowance without leaving the conversation).
- What it is NOT: it never returns a trade instruction, it is not a forecast service, it is not investment advice, and it is independent research — not affiliated with or endorsed by Al Brooks.

## Pages

- [Home](https://maddog.finance/): what MADDOG does, the historical replay demo, pricing, and the free one-read form.
- [About MADDOG](https://maddog.finance/about): what MADDOG is, who it is for, how accurate it is, what it costs, and what it is not — the canonical answer page.
- [Sample reads](https://maddog.finance/sample): real opening-read emails written by the exact production pipeline on real bars (NVDA, TSLA, SPY, BTC-USD).
- [API reference](https://maddog.finance/docs): endpoints, symbol formats, window parameter (rth / asia / gap), and what every response field means.
- [Blog](https://maddog.finance/blog): measured price-action statistics with archived results, reconstructions, pre-registered tests and their limits labelled separately.
- [Opening range breakout strategy: 16 years of ES data](https://maddog.finance/blog/opening-range-breakout): separates an archived 18-bar opening-range result (83.5%, n=3,578) from MADDOG's live rolling 10-bar event. The current rolling RTH references are 52.1% within 5 bars (n=3,925), 62.6% within 10 (n=3,922), and 70.6% within 20 (n=3,915). These are different experiments and none of the figures is a trading-strategy win rate.
- [Opening range breakout failure rate](https://maddog.finance/blog/opening-range-breakout-failure-rate): the exact prior-10-bar, close-based definition behind the current 52.1% / 62.6% / 70.6% cumulative ES failure references. These are event outcomes, not a strategy loss rate or fade win rate.
- [When does the high or low of the day form?](https://maddog.finance/blog/when-does-high-or-low-of-day-form): at least one final ES session extreme had appeared in 26.3% of archived sessions after 5 minutes, 58.3% after 35 and 81.2% after 90. The page keeps archived and reconstructed conditional samples separate and explains why this is not a directional forecast.
- [Opening range breakout methodology](https://maddog.finance/methodology): exact study definitions, data scope, reconstruction limits and current API references. The original conditional model was pre-specified; the published AUC 0.646 / ECE 5.1% / n=2,637 tuple is the later production ADR5-port validation on NQ 2023+.
- [ETH vs RTH: which session should you chart?](https://maddog.finance/blog/eth-vs-rth): only 40.4% of QQQ's five-minute bars fall inside regular hours (09:30-16:00 ET), a regular-session bar travels about 2x as far as an extended-hours one, and on 93.2% of days the day's high or low printed outside regular hours (84.7% using bar closes only). Measured on 59 days, 2026-05-22 to 2026-08-17.
- [Pulse — 24-hour movers](https://maddog.finance/pulse): a full-market scan (312 instruments in the 2026-08-19 run) narrowed to the ten biggest 24-hour moves, in five languages (zh/en/ja/ko/es, one URL per language via ?lang=). Every row carries the measured fields — change, intraday range, volume versus the median of the prior 20 bars, sigma against the instrument's own 30-day realised volatility, rank among the past 100 complete sessions — plus the endpoint each number was measured from. Two measurement bases are deliberately different and both are stated on the page: crypto uses the exchange's rolling 24 hours, US equities use previous close to last price. Attribution is gated, not narrated, and every attribution carries its evidence tier on the badge. A first-party, machine-readable event (an SEC filing, an FDA or Treasury release) is the strongest grade and the ONLY thing a published number may come from; press coverage may carry an attribution too, but it is badged secondary and its confidence is capped at medium, because a report tells us what happened without telling us it happened from the party it happened to. Either way the event must land BEFORE the volume break, be within 12 hours of it, and be in scope for that instrument (issuer news must name the instrument; market-wide news only applies to assets that are themselves macro-priced; a filing may only vouch for a multi-ticker story if its own text mentions that story's subject; articles that merely explain a move that already happened — "Why Did X Jump Today?" — are demoted to leads). When the news broke premarket and the instrument gapped at the opening bell, the lag is not measurable and the row prints no lag at all rather than a number measured from the bell. Rows where nothing clears those gates are labelled unattributed and left blank; the count varies run to run, and data.json is rewritten on every scan, so read the badges on the page rather than any number quoted here. Each run is also frozen at a permanent per-day address — https://maddog.finance/pulse/YYYY-MM-DD, one page per UTC date, which never changes once published; every archived date is listed at https://maddog.finance/pulse/archive.
- [Scoreboard](https://maddog.finance/scoreboard): the public forward record. Every US trading day at 11:00 AM ET the day-session read for ES=F, NQ=F and QQQ is snapshotted verbatim; after the close the row is scored on whether a new session high or low printed after the read, against the extension rates the read carried. Ties count as "no". Recording began 2026-08-25; rows are never edited and days with no live read stay visible as gaps. The raw feed is free and keyless at https://api.maddog.finance/v1/scoreboard. The page also carries a clearly-labelled hindsight replay (frozen weights on Yahoo's 60-day history, sessions strictly before 2026-08-26) for format demonstration and calibration cross-check — it is not a track record and never mixes with the forward rows.
- [Interactive API spec](https://api.maddog.finance/docs): always in sync with the running service.

## Contact

- info@maddog.finance