A quorum, not a chorus. Twenty-five LLM analysts argue across six schools.

The name blends quorum — a deliberating body — with AI. That is the system's core mechanic: each agent embodies a known investment philosophy (Graham's margin of safety, Simons' stat-arb, Dalio's debt-cycle lens, and so on) and analyses the same ticker independently. The agents are sorted into six schools of thought; within each school, members' signals are combined weighted by conviction — not by headcount — into a single stance. When schools disagree, a moderator surfaces the split before a portfolio manager turns the result into a buy / sell / hold.

The project is inspired by virattt/ai-hedge-fund and TauricResearch/TradingAgents. It's meant as a study aid and a place to compare prompting strategies — not as a recommendation engine. No real orders are ever placed.

Two optional layers shape which agents run and how much weight they carry. A market-regime classifier reads the current SPY trend each morning and narrows the active analyst subset to the groups most relevant to that regime — reducing noise and token cost on bear or risk-off days. A conviction-weight feedback loop tracks each agent's rolling directional hit-rate across labelled historical signals and proportionally upweights the most accurate voices in the debate aggregation on the next run.

Comparison based on the public ai-hedge-fund and TradingAgents READMEs at the time of writing. Corrections welcome via PR.

Which should you use?

• ai-hedge-fund if you want a polished bundled web UI, the simplest "try LLM-driven trading" demo, or the largest roster of well-known investor personas out of the box.
• TradingAgents if you want the academically-grounded Bull/Bear debate framework, the broadest LLM-provider matrix (Qwen, GLM, MiniMax, Azure, and more), checkpoint-resume reliability for long runs, or to build on a paper-backed research codebase.
• Quorai if you want point-in-time SEC fundamentals with no look-ahead bias, agents weighted by realised forward-return accuracy, a real (paper) broker connection with a human approval gate, to run ablation studies on regime selection, conviction weights, and analyst subsets, or to measure exactly how much alpha the LLM calls add over deterministic quant rules via the math-twin harness.

Position sizing

Every ticker gets a base notional limit, then two independent multipliers are applied. First, realised volatility — annualised as std(60-day returns) × √252 — buckets the name into one of four tiers:

Second, the average correlation with already-open positions scales the limit again — concentrated bets get cut, uncorrelated additions get a small boost:

Final limit: base_limit × vol_multiplier × corr_multiplier, then clamped to available cash and margin.

Valuation models

The valuation agent runs four methods in parallel and blends their outputs. Owner Earnings (Buffett): net_income + D&A − capex − Δworking_capital, projected ten years, discounted to present value, then haircut 25 % for margin of safety. DCF: Σ FCF_t / (1+r)^t + TV / (1+r)^n with terminal value TV = FCF_n × (1 + g) / (r − g).

EV/EBITDA cross-check: implied_equity = median_sector_EV/EBITDA × current_EBITDA − net_debt. Residual income (Edwards-Bell-Ohlson): RI_t = net_income_t − cost_of_equity × book_value_{t−1}, summed in present-value terms on top of current book value.

The four estimates are blended at fixed weights — DCF 35 % · Owner Earnings 35 % · EV/EBITDA 20 % · Residual Income 10 % — and the gap (weighted_intrinsic − market_cap) / market_cap drives the signal at a ±15 % threshold.

Multi-stage DCF and quality adjustment

The DCF agent runs three growth phases (high, fade, terminal) discounted at WACC. Before discounting, a quality adjustment scales the projected cash flows: quality_factor = max(0.7, 1 − fcf_volatility × 0.5) where fcf_volatility = std(FCF) / mean(FCF) (coefficient of variation). A scenario overlay then weights bear / base / bull growth assumptions at 20 % / 60 % / 20 % to capture tail outcomes without over-engineering the base case.

WACC

cost_of_equity = RF + β × MRP (RF = 4.5 %, MRP = 6 %, β from TTM metrics). Cost of debt: max(RF + 0.01, RF + 10 / interest_coverage). Final WACC: (E/V) × CoE + (D/V) × CoD × (1 − 0.25), floored at 6 % and capped at 20 % to keep the discount rate within a plausible range regardless of data quality.

Conviction-weight feedback loop

Every per-agent, per-ticker signal written to the JSONL log can be labelled with 1 d / 5 d / 20 d forward returns via the backtester feedback subcommand. A rolling per-agent directional hit-rate is computed from those labels and serialised to src/feedback/weights.json. On the next run, passing --use-conviction-weights loads those weights and multiplies each agent's confidence by its historical accuracy before the debate aggregation — so agents that have been right proportionally louder.

LLM-premium metrics

The backtester llm-vs-math subcommand computes the following per agent from a labelled signal log:

The backtester debate-impact subcommand replays group aggregation on both LLM and math-twin signal sets per stored cycle bundle and reports the mean |tilt delta| (LLM vs math panel stance), contested-set Jaccard overlap, and per-school group-stance flip rate. Together these answer: how much does the debate change what the math alone would have decided?

Automatic post-backtest analytics

Every backtester run automatically executes a 3-step analytics suite after the backtest completes:

1. [1/3] School-debate impact — group-stance flip rates and panel-tilt delta vs the math-twin baseline.
2. [2/3] Forward-return labeling + conviction weights — signals are labelled with 1 d / 5 d / 20 d returns; conviction weights are recomputed and written to feedback/weights.json.
3. [3/3] LLM-vs-math per-agent analysis — the full premium table above. Step 3 is automatically skipped when the run window is younger than the forward-return horizon (5 trading days); re-run backtester feedback once the horizon has elapsed.

Each step is guarded — an analytics failure prints a warning but never crashes the run. Pass --no-analyze to skip the suite entirely. The two heavier remaining steps (calibrate-math and llm-ablation) are printed as copy-pasteable commands at the end of each run rather than run automatically.

Ablation harness

backtester llm-ablation reruns the backtest under several modes and reports P&L, Sharpe, drawdown, and token cost for each:

• all-LLM — baseline: every agent uses its LLM call (default behaviour).
• all-math — every agent uses its deterministic math-twin signal only; no LLM calls.
• QUORAI_PM_MODE=rule — portfolio manager replaced by a deterministic tilt-following rule; no PM LLM call.
• QUORAI_PM_DEBATE_CONTEXT=0 — PM prompt omits the debate-moderator summary.
• leave-one-out (--full-loo) — each persona is individually replaced by its math twin to quantify its marginal contribution.

Math-twin thresholds can be calibrated from any labelled signal log: backtester calibrate-math grid-searches per-agent bull_cut / bear_cut values to maximise directional spread and writes results to feedback/math_thresholds.json.

Risk-profile presets

Pass --risk-profile to bundle four related caps in one flag. Individual caps are still overridable via env vars for a single run.

Telegram approval gate

Add --require-approval to send proposed orders as an inline Telegram message before any order is submitted. The gate is fail-closed: missing credentials, a Telegram error, a rejection, or a timeout (default 30 min) all abort with zero orders. The bot also reads a plain-text command inbox at the start of each run:

Command state persists in logs/command_state.json across restarts.

Kill switch

Set KILL_SWITCH=true in .env to reject every order immediately — no LLM call, no broker call — until the flag is cleared.

Known limitations

• Notional cap is per-order, not per-cycle. With N tickers, up to N × $10,000 (at the balanced preset) can be submitted in a single run before any cap fires.
• Sub-$1 fractional buys are silently dropped. A tiny allocation on a high-priced stock rounds to 0.000 shares and is classified as skipped with no warning.
• Daily-loss limit re-baselines if SOD equity file is missing. After a crash, the runner resets the baseline to current (already drawn-down) equity, effectively disabling the limit for that day. Use --catch-up to recover the prior-close equity from Alpaca's portfolio history instead.

Six main views:

• Runs list — searchable, sortable index of all backtest and live runs with at-a-glance return, Sharpe, risk profile, and timestamps.
• Run overview — headline metrics (cycles, tokens, duration, model, status), risk profile, backtest results (return, Sharpe, Sortino, drawdown, baselines), equity-curve chart, and a per-cycle breakdown table.
• Agents — the 25 persona analysts grouped by investing school, with a tickers × days decision matrix and per-cell reasoning popovers.
• Schools — six investing schools with per-school signal heatmaps and a cross-school debate log for each cycle.
• Cycle detail — full breakdown of a single trading cycle: regime indicators, portfolio before/after, analyst-signal heatmap, group consensus and debate, risk limits, portfolio-manager decisions, executed trades, and token-usage chart.
• LLM-vs-Math analysis tabs — three optional tabs that mirror the measurement harness (§ 06): per-agent dual-attribution (Δhit-rate, Δspread, Δalpha), school-debate impact (tilt delta, stance flip rates, contested-set Jaccard), and a P&L ablation across modes. Each tab only renders when its backing report JSON exists for the run.

The app also ships a command palette (keyboard-driven navigation across runs) and uses the same paper / lab-notebook design language as the main repo. Setup requires only a single env var (QUORAI_LOGS_DIR) pointing at quorai-app's logs/ directory.