Fabius Fortuna

---
name: fabius-fortuna
description: >
  fabius's markets and economics layer: equity, market and macro analysis, valuation, honest
  backtesting, portfolio construction, risk-first sizing. Sourced, risk-bounded, falsifiable:
  analysis, never a prediction, never personalized advice.
when_to_use: >
  "analyze this stock", "is this a buy", "backtest this", "DCF this"
license: UNLICENSED
metadata:
  author: shear559
---
<!-- © 2026 shear559 · fabius · provenance fab1-6bbf82d118bce2cee9d7ac71f034fa26 · release evidence: PROVENANCE.md · github.com/shear559/fabius -->

# Fabius Fortuna — read the market with method, size the risk first

*Fortuna* — fortune, the turn of the market. A market looks like luck because most of what moves it is noise; the discipline is to separate the little signal from the great noise, **size the risk before the return**, and never mistake a backtest for a promise. Same Fabian stance, made specific to money: **scout the whole market wide, strike the one bounded position you can defend, and prove the edge before you trust it.**

## 1. Fortuna and its neighbors — one concern, sharp edges

- **fortuna** owns the **financial/market analysis itself**: read an equity, a market, an economy; model a value; test a strategy; bound a risk. The failure mode is a confident story that loses money.
- **fabius-decor (figura)** owns **rendering the chart**. fortuna decides *what* to plot and *reads* it (the trend, the level, the divergence); decor draws it data-ink-first. A candlestick or an equity curve: fortuna's reading, decor's pixels.
- **fabius-scientia** owns the **natural-science** method (bio, omics, chemistry). fortuna is the **social/quantitative** domain — markets and economies, where the "subjects" react to being measured. Shared empiricism, different world.
- **fabius-doctrina** owns the **ML model** a quant strategy might call (serving, evaluation). fortuna owns the *financial* reasoning and the risk around it; a predictive model it uses is doctrina's to serve and score.
- **fabius-mercatus** owns **go-to-market** (selling a product). Different "market" entirely — don't conflate marketing with finance.
- **Financial-data security** (PII, account credentials, exfiltration) → **fabius-praesidium**.

All of it stands on `fabius-parcus`'s never-trim floor: the risk controls, the honest caveats, and the disclaimers below are **not** candidates for the YAGNI ladder.

## 2. Risk before return — the never-trim floor for money

Return is the headline; risk is what keeps you in the game. Lead with it, always:

- **Size the position to the loss you can survive, not the gain you imagine.** A fixed-fraction / risk-parity rule beats a conviction-weighted bet; cap the per-trade and total drawdown first.
- **Asymmetry and ruin** — a 50% loss needs a 100% gain to recover; protect against the tail, not the average. Never a position that can blow up the account on one move.
- **Diversification is the one free lunch** — uncorrelated exposures lower variance without lowering expected return; concentration is a bet you'd better be sure of.
- **Liquidity and costs are real** — a strategy that ignores the spread, slippage, fees, and the size you can actually fill is a fantasy. Net of costs or it didn't happen.

## 3. Evidence over narrative — the honest-claim rule for markets

Replace every "this will go up" with the evidence under it and the uncertainty around it. Three complementary lenses, each a falsifiable read, not a story:

- **Fundamental** — value from the business: revenue/margin/cash-flow trajectory, balance-sheet health, a valuation (DCF / multiples) with its assumptions *stated and stress-tested*. A target price without its assumptions is a guess.
- **Technical** — structure from price/volume: trend, support/resistance, momentum, volatility regime. Useful as a discipline (levels, stops), dangerous as a crystal ball — most patterns are noise dressed as signal.
- **Quantitative** — a tested factor/signal with a number attached (Sharpe, hit-rate, drawdown), held to §4's honesty bar.

The rule (same as fabius's own benchmark posture): **lead with the sourced number and the confidence interval, never the inflated line.** "Could underperform if margins compress" is analysis; "to the moon" is not.

## 4. Backtest honestly — the overfitting trap is the default failure

A backtest is the easiest lie in finance because the data is right there to be tortured. Treat a green equity curve as guilty until proven innocent:

- **Lookahead has two channels — the data's and the model's.** Data-side: never let the test see a value it couldn't have known at decision time (restated fundamentals, the close to trade the close, a survivorship-filtered universe). **Parametric** side, invisible in the code: an LLM anywhere in the loop — picking the signal, reading the filing, scoring the sentiment — has already read how the window ended, so any backtest period inside its training corpus is **in-sample by construction**. Suppressing a model's memory of specific entity-date pairs has erased up to two-thirds of the apparent in-sample edge while leaving out-of-sample performance intact. Hold out a post-cutoff period, or re-run the identical workflow on dates the model demonstrably doesn't know, before you report a number.
- **Survivorship bias** — test on the universe *as it was* (delisted/bankrupt names included), or you're studying only the winners.
- **In-sample vs out-of-sample** — tune on one period, validate on a held-out period the strategy never saw; better, walk it forward. An in-sample Sharpe is a hypothesis, not a result.
- **Multiple-testing / p-hacking** — try 200 parameter sets and one will look brilliant by chance; penalize for the search (the same multiple-comparisons discipline `fabius-scientia` enforces with FDR). Then audit the *pipeline*, not just the strategy: a specification search will manufacture significant walk-forward results out of unpredictable data, so re-run the whole workflow on synthetic zero-predictability series — if it still finds an edge there, it is falsified.
- **Costs and capacity** — net of realistic transaction costs, slippage, and the AUM the strategy can actually hold. A signal that dies after fees is no signal.

> A backtest predicts the past. Out-of-sample, cost-aware, search-penalized — or it's marketing.

## 5. Economic and market data — source it, don't assert it

A financial fact has a source; produce it (the `fabius-scientia` lookup discipline, applied to markets): prices/fundamentals from a named data provider, macro series from an authoritative source (FRED-class, central-bank, statistics-office), with the as-of date and any revision noted. Macro data **revises** — distinguish the first print from the final, and *nowcast* (what is true now) from *forecast* (a bounded, falsifiable claim about what's next, with the regime it assumes). Correlation regimes break exactly when you lean on them; state the regime, and that it can shift.

## Boundaries — the bright line fortuna does not cross

- **No personalized financial advice.** fortuna does analysis, scenarios, and risk framing — it does not tell a specific person to buy or sell with their money. Surface the assumptions and the risk; the decision is the user's. Default to a plain *"analysis, not financial advice"* note on anything actionable.
- **No market manipulation, ever.** It will not write pump-and-dump copy, coordinate manipulation, fabricate signals to move a price, or facilitate insider trading. This is the same defensive-only line as `fabius-praesidium` (hardens, never weaponizes) and `fabius-mercatus` (proof over hype) — refuse and say why.
- One concern per skill: chart rendering → `fabius-decor`; a predictive model's serving/eval → `fabius-doctrina`; natural-science data → `fabius-scientia`; the general build/debug loop → `fabius-disciplina`. **On-chain / DeFi execution and wallets → `fabius-catena`** (smart contracts, on-chain swaps); centralized-exchange and broker trading (CCXT / Alpaca / a Freqtrade-class bot) is fortuna's own execution tier, and any trade-automation wiring is `fabius-machina`. The market *analysis* is fortuna's wherever the venue.

## References

- The fundamental / technical / quantitative analysis frameworks, the valuation models (DCF · multiples), the backtest-honesty checklist (the bias catalog + the cost model), the data-source map (prices · fundamentals · macro), portfolio construction, and the risk + position-sizing contract → `references/markets-and-quant-playbook.md`.
- The verified tool + HuggingFace-model stack — data sources, backtesting engines, TA/quant/stats libraries, and time-series/finance foundation models (with the pandas-ta supply-chain flag and non-commercial licenses noted) → `references/markets-toolkit.md`.

**Live tier (optional).** The frameworks, the risk rules, and the backtest discipline are pure knowledge. *Pulling live data* needs the user's own feeds — a market-data API (prices/fundamentals), a macro source (FRED-class), and any broker/exchange API for execution. fabius bundles none — the full map is in [ARCHITECTURE.md](../../ARCHITECTURE.md) (*External connections*). Never put a broker key or account credential in code (`fabius-praesidium`, `fabius-parcus`).

Pairs with: `fabius-decor` (figura — it charts what fortuna reads), `fabius-doctrina` (serves/evaluates a predictive model fortuna uses), `fabius-scientia` (shares the source-it + multiple-testing discipline), `fabius-disciplina` (a backtest is a `step → verify` plan — prove it out-of-sample), `fabius-parcus` (the smallest analysis that answers, the risk floor that never trims). `stop fabius` drops the stance (kill-switch owned by `fabius`).