MonarqiLabs — Stress Trading Desk for IDX Stocks

BBCA 8,725 ▲0.9% BBRI 5,140 ▼1.2% BMRI 7,050 ▲0.3% TLKM 3,180 ▲0.4% ASII 5,625 ▼0.6% BBNI 5,475 ▲1.1% ANTM 1,685 ▲3.2% MDKA 2,450 ▼1.7% ICBP 11,300 ▼0.3% UNVR 2,342 ▲0.8% ADRO 2,610 ▲1.4% PTBA 3,120 ▼2.1% GOTO 68 ▲2.3% BRIS 2,890 ▼0.9% KLBF 1,310 ▲0.2%

01 · The Premise

Two worlds, one seed.

Every experiment runs twice from the same seed: a scenario world that receives your shock, and a baseline world that lives without it. Both trade the same 961 stocks with the same ten agents.

The delta between the curves is the experiment: price impact in basis points, days to recovery, and which agent moved first.

ΔImpactshock-day price move in the scenario world, measured in bps against baseline

t+Overshootdoes the market undershoot fair value before turning, and by how far

t→Recoverydays until the scenario index rejoins the baseline path

ΣAttributionwhich personas bought the dip, which froze, which made it worse

02 · One Trading Day

A day runs in seven moves.

The engine is a strict tick loop. Rules think every day, which keeps runs cheap and deterministic. The LLM thinks only when something happens: a shock lands, volatility doubles, or an account crosses a psychological line. Reflections are cached, so a rerun of the same seed costs nothing.

01Fundamentals driftevery stock's true_value takes its daily random walk with mean reversion

02News layeroptional Tier-2 news items are emitted; News Nelly trades on sentiment alone

03Shock injectionif one is scheduled today, it lands. One shock per day, never two

04Agents computeall ten personas price their orders from Tier-1 rules + memory notes

05Matchingorders meet the book, prices update, fills are recorded

06Vol checkif realized vol exceeds 2× its 20-day mean, affected agents reflect (Tier 2)

07Log & emitthe day is appended to events.jsonl, the single source of truth

03 · The Roster

Ten desks. Ten temperaments.

Ten different nervous systems, not ten copies of one strategy. Each keeps an append-only memory log, and Tier-1 rules read yesterday's notes as today's modifier.

00Value VeraBuys below 0.85× intrinsic, sells above 1.15×slow · conviction

01Momentum Max20-day MA cross, momentum scoringfast · signal-chasing

02MeanRev MiaZ-score reversion, ±1.5σ bandfast · mean-centered

03FOMO FredChases green: buys after back-to-back up daysimpulsive · herding

04Panic PeteExits past 8% drawdown, holds ≥60% cash when scaredemotional · loss-averse

05News NellyTrades the news sentiment score, ignores pricereactive · tier 2

06Contrarian CarlFades short-squeezed rallies and panic washoutsslow · contrarian

07Passive PatEqual-weight buy-and-hold, places no active ordersno memory

08Fund FMVol-scaled sizing, stop-loss, diversification capsdisciplined · systematic

09Macro MargePositions by the rate proxy: longs when rates fallevent-driven

Control group Agent 07 never trades. An equal-weight buy-and-hold runs in every experiment, because a claim that active management beat the market means nothing without a baseline.

04 · Shock Injection

One shock at a time.

Shocks are written in the same grammar the desk uses: day, type, ticker, severity, one per line, one per day. An earnings miss lands on one name; a rate cut hits the whole board. The LLM narrates each shock into the news rail, then every affected agent reflects on it and adjusts tomorrow's behavior.

earnings_miss rate_cut margin_call analyst_downgrade short_squeeze index_rebalance

shocks.txt, same grammar as the desk

# one per line: day|type|ticker|severity

150|earnings_miss|BBRI|0.80

500|rate_cut|global|0.50

740|margin_call|ANTM|0.65

# day multipliers, optional global drag

501|mult|0.75

Severity 0.80 is not decoration. It scales the impulse the price engine absorbs that day, and it decides how many agents trip a Tier-2 reflection.

05 · The Lab Keeps Score

Watchability is the gate. Statistics are the output.

Every metric is computed per agent, per run, then averaged across a batch of 50 seeds with confidence intervals. A single run is a story; fifty runs are a result.

Total P&Lper agent, per run, averaged across seeds with CI+12.4%

Sharpe ratiofrom daily returns1.31

Max drawdowndeepest peak-to-trough equity loss−18.2%

Shock reactionprice impact in bps, days to recovery, and which agent moved first−212 bps · 41d

Crowd rationality1 − (excess volume + price overshoot + panic-sell volume); a declining trend means the market is getting stupid0.87

Strategy driftcorrelation between an agent's actual decisions and its stated strategy, over time0.93 → 0.71

Turnovervolume traded vs average portfolio value per day0.42×

06 · Event Sourcing

Every run is a replayable truth.

The engine never mutates the log. The UI and the metrics both read from it. Replay any finished run and scrub to any day; live mode and replay mode are the same data through the same pipe.

Deterministic by construction: the same seed produces the identical market, and LLM reflections are cached by seed, day, agent, and trigger, so a scientific rerun costs nothing.

experiments/ml_001/events.jsonl

{ "event": "shock", "day": 150, "type": "earnings_miss", "stock": "BBRI", "severity": 0.8 }

{ "event": "order", "day": 150, "agent": "panic_pete", "stock": "BBRI", "side": "sell", "qty": 400, "reason": "tightened_stops" }

{ "event": "fill", "day": 150, "agent": "panic_pete", "stock": "BBRI", "price": 8.62 }

{ "event": "reflection", "day": 150, "agent": "panic_pete", "trigger": "vol_spike", "note": "Market breaking down; tightening stops to 3%.", "confidence": 0.6 }

{ "event": "reflection", "day": 151, "agent": "contrarian_carl", "trigger": "position_breach", "note": "Washout pricing; starting a small long.", "confidence": 0.55 }

{ "event": "day_end", "day": 151 }

The desk is live

Run the seed. Stress the market.

Open the trading desk, write one line of shock, and watch ten agents argue with a market about what things are worth.

Open the Desk → How it works

Deterministic · seed 42 · 10 agents · 961 tickers