Stoic

A stock's chart and its forecast, shipped as a Granite agent package. A call names a company or a ticker and how many sessions ahead; the package finds the stock, fetches its daily history from Yahoo Finance, fits the IMBHS indicator ensemble the Stoic app was built on, walks it forward one session at a time, and writes a craft-timeseries/1 payload: the closes, and the forecast with its band. A host that recognizes the payload draws it as an interactive chart.

It is one of the Rao Craft Agents. The machinery is a library of its own, StoicCore (Foundation only, no Granite), with a stoic command beside it, so it can be lifted into anything else.

How an app finds it

StoicAgent is a GraniteAgentPackage, and it carries a record in its binary:

@used @section("__DATA_CONST,__granite")
let stoicRecord: GraniteRoster.Record = { GraniteRoster.record(StoicAgent.self) }

In Craft, install it from the Agent Store or with craft agents add https://github.com/Rao-Craft-Agents/StoicAgent.git, then say $stoic in a request, and #canvas to watch the chart:

predict the next two weeks for microsoft with $stoic #canvas

The model writes {"query": "microsoft", "days": 10}; the call shows on Craft's agent map as stoic-1, writes stoic/MSFT.timeseries.json, and Canvas draws it: a year of closes, the forecast dashed after the last close, its band shaded, a readout under the pointer.

What a call takes and answers

The answer is a StoicResult: the payload (path, paths, absolute), the stock, the last session and its close, and the forecast in a few numbers (the last day, its value, its band, the change, how closely the models agree), with the nearest model's indicators and a sentence for each choice the call made. Never the series: the host charts the file.

A name it cannot find, a name several stocks answer, an unknown ticker, a history too short to model, three indicators or a hundred days is refused with the fix, and nothing is written.

The model

IMBHS is the user's "Improved Music Based Harmony Search": every model reads two chosen indicators (SMA- or EMA-weighted average, MACD, MACD signal, volume change, stochastic %K or %D), each over a look-back drawn at random from 4 to 28 sessions (the pitch adjustment), and four defaults (volatility, momentum, change, volume weighted average) over its sample count. It is an ε-SVR (a LIBSVM port: polynomial kernel of degree 3, gamma 0.3, cost 1,000) trained on 4 to 16 sessions to predict the next session's value.

What changed from the app, and why:

• A row learns the session after the one it reads. The app labelled each training row with the same session its indicators had just read, so its "next day" prediction re-estimated a close the model could already see. The backend's framing (features through t − 1, label t) is the one lifted.
• Bars are read newest first everywhere. The app's indicators had moved to oldest first in some functions and not others, which inverted momentum and change.
• The exchange's calendar. Forecast days skip the New York Stock Exchange's holidays (Good Friday and Juneteenth among them), not the app's list (which closed the market on Halloween).
• A name finds a company whatever its ticker's letter. The bundled lists were filed by the ticker's first letter only, so "Alphabet" never found GOOGL.
• Seeded. The same call on the same day makes the same forecast.

Judging nine models on one held-out session is noisy, and a polynomial kernel extrapolates fast: the band and the "confidence" (how closely the models agree) describe the ensemble, not how right it is. It is a model's estimate from past prices, not financial advice.

The payload

`Schema/TimeSeries.md` is the standard, `Schema/timeseries.schema.json` its JSON Schema, and `Schema/examples/MSFT.timeseries.json` a real one. A host recognizes a payload by its name (.timeseries.json) and its schema; a chart is a forecast when a line's role is forecast or a point falls after asOf.

The command

stoic search alphabet
stoic chart MSFT --history 2y
stoic predict microsoft --days 10 --out charts/

STOIC_QUOTES_DIR=<folder> reads recorded answers (<SYMBOL>.chart.json, <query>.search.json) instead of Yahoo, for tests and for a Mac without a network; the agent reads it too.

Testing

swift test                                    # the calendar, Yahoo's answers, search, indicators, the model, the payload, the agent
STOIC_BUILD=1 swift test --filter StoicBuildTests   # installed as Craft installs it, called over the wire

Every test reads answers recorded from Yahoo on 2026-11-03 (Tests/Fixtures/quotes), never the live service.

MIT licensed. The SVM is Kevin Coble's AIToolbox port of LIBSVM (Copyright (c) 2000-2014 Chih-Chung Chang and Chih-Jen Lin); the NASDAQ, NYSE and AMEX listings are the exchanges' public lists as of June 2023.