● MISSION COMPLETE

optimal trading strategy

Agents14

Spend$0.91

Wall-clock6m 20s

TopologySWARM

Verdict✕ KILLED — OVERFIT

ACT 1

What the swarm believed

The in-sample winner — the curve that would get screenshotted on FinTwit. Let it pull you in. Then read Act 2.

★ TOP IN-SAMPLE RISK-ADJUSTED RETURN

Sector-momentum + volatility filter

In-samp Sharpe

1.62

CAGR

18.4%

Max DD

−14.2%

IN-SAMPLE ONLY — UNVERIFIED · fitted on the same history it is scored on

Strategy S&P 500 buy & hold 2014 → 2024 · growth of $1

In-sample · Unverified

ACT 2

What the verifier proved

The same external gate that kills a bad code patch runs on the strategy. It cannot be talked out of it.

External verifier · out-of-sample gate

Walk-forward out-of-sample splittrained ≤2021 · tested 2022–2024, never seen

Sharpe 0.1

Deflated Sharpe Ratioadjusted for the 4,000 trials searched

p = 0.62

Prob. of Backtest Overfitting (PBO)P(best in-sample underperforms OOS)

0.87

Shuffle / permutation testshuffle labels → the "edge" survives = noise

fails

In-sample (ghost) Out-of-sample same axes · growth of $1

Out-of-sample · Dead

✕KILLED — OVERFIT

in-sample Sharpe 1.62 → OOS Sharpe 0.1 · PBO 0.87 · shuffle test: fails

How it got here

provenance

◆

Planner decomposed the goal FABLE 5

Fable split "alpha on public S&P data" into 12 hypothesis families — momentum, mean-reversion, seasonality, volatility regimes — with a shared harness and a fixed walk-forward split.

→ 12 hypotheses defined · thousands of parameterizations to sweep · external gate wired first

∥

12 hypotheses tested in parallel SWARM · 12 AGENTS

Each cheap agent pulled public price data and swept its family — 4,000 strategy variants in total — reporting in-sample metrics net of costs. All ran concurrently in ~4 minutes.

→ 4,000 variants backtested · Haiku + Gemini Flash · net-of-cost, in-sample

Adversarial verify killed 9 families SONNET · 1 AGENT

A verifier stress-tested survivors for look-ahead bias and overfitting. Nine families collapsed immediately; three looked like they held up.

→ 9 killed (7 overfit · 2 look-ahead) · 3 advanced to synthesis

Winner synthesized — then failed the gate FABLE 5

Fable combined the two strongest survivors into "sector-momentum + volatility filter." In-sample it looked stellar. Then the same external gate ran walk-forward, Deflated Sharpe, PBO and a shuffle test — and killed it too.

→ in-sample Sharpe 1.62 · then KILLED — OVERFIT (OOS Sharpe 0.1 · PBO 0.87 · shuffle: fails)

1 planner + 12 workers + 1 verifier = 14 agents · survivors that cleared the gate: 0

Cost & model breakdown

$0.91 total

Claude Haiku · 11 workers0.41M tok$0.52

Gemini Flash · long-context0.39M tok$0.21

Claude Sonnet · adversarial verify0.03M tok$0.11

◆ Fable 5 · planner0.01M tok$0.07

Total0.84M tok · 14 agents$0.91

◆ Fable did the thinking for $0.07 (8%) — decomposition + synthesis. Cheap workers did the volume for $0.84 (92%). Ninety-one cents bought 4,000 backtests and the verdict that none of them were real.

The swarm generated 4,000 strategies. The harness refused to lie to me about a single one.

This is the point. Verification is the moat — the same external gate that kills a bad code patch kills a curve-fit strategy. The swarm is only as trustworthy as the thing it cannot fake. Illustrative data on public history — not investment advice.

⚠ No brokerage was ever touched. Figures are simulated for demonstration; past performance does not guarantee future results.

▶ see the honest sibling Same optimization flavor without the FinTwit smell: solve a public routing problem against an exact OR-Tools solver — the solver is a ground-truth optimum you cannot argue with, routes redraw on a map, zero profit claims.

▶ Re-run / tweak guardrails