QuantumMind

Hero scenarios: swarm size:

Expert Layer

Bull vs. Bear debate

Swarm (0)

drag to orbit · scroll to zoom · click a dot

Synthesis

By cohort

Directional tracking

Polls yfinance news for each ticker, runs the pipeline on fresh headlines, streams results.

auto-refresh (5 min)

Watchlist:

Compare two scenarios

Preset pairs for the pitch, or pick your own below.

The cross-pod problem

A $1B long/short fund organizes its PMs into sector pods. The tech PM doesn't talk to the energy PM. When a catalyst hits, the obvious single-stock trade is already priced in by the next session. The alpha lives in the cross-sector ripple — second-order effects that diffuse over a week as different desks react in turn.

QuantumMind reads every catalyst, generates the cross-sector trade chain automatically, and surfaces it to the CIO. Below is what happened when we ran this on real news for the past 8 weeks.

Cross-pod chain backtest · live news events · no curation

Equal-weighted portfolio, daily mark-to-market, 10bps roundtrip costs. Sharpe = non-overlapping weekly returns × √52 (autocorrelation-corrected).

iteration:

Portfolio annualized

—

Weekly Sharpe

—

Sharpe ratio (annualized)

Hit rate

—

Bootstrap P(positive)

—

10k resamples · seed=42

Mean per trade (net)

—

Max drawdown

—

Sample window

—

Stress test · re-roll the LLM N times, what's the worst case?

The shipped chains.json was one draw from the LLM. We re-rolled the model from scratch N× on the same 200 events. These are the fresh-roll outcomes.

view	metric	worst	median	best	std

Reproducibility: bootstrap seed=42, prices cached. Each row = one metric across all N regenerations. Run scripts/run_chain_stress_test.py 5 to recompute (each re-roll is ~5 min of LLM calls).

Methodology + reproducibility ↓

Pipeline. Each news event triggers one LLM call to QuantumMind AI with a cross-sector reasoning prompt. The model returns 3–5 trade legs (sector, ticker, direction, rationale). Each leg held 5 trading days.

Execution. VWAP entry on next trading day after news, VWAP exit 5 days later, 10bps roundtrip costs. yfinance OHLCV. No look-ahead.

Sharpe. Daily portfolio returns are auto-correlated when positions overlap, which inflates daily Sharpe. We bucket into non-overlapping 5-day windows and annualize × √52 — that's the number shown above. Daily Sharpe (the inflated one) is also computed for transparency, available in the JSON.

Bootstrap. 10,000 resamples of trades with replacement, seeded with 42 → reproducible. P(positive) = fraction of bootstrap means greater than zero.

Reproducibility. The chains.json is frozen; same chains + same yfinance prices + same seed = bit-identical numbers. Run scripts/chain_diagnostics.py any time to verify. To regenerate chains, delete chains.json and run scripts/run_chains.py (this DOES involve LLM calls and is non-deterministic — that's the only place randomness enters the pipeline).

Iterations shown above. v1 (mixed-horizon) is the baseline before constraint. v2 locks the prompt to 1-week. v3 additionally filters to LONG-only because the model has clear edge going long but bleeds value going short. We deploy v3.