Engagement bait: "results are INCREDIBLE:" is followed by the paper's first page, and no P&L appears in the post, article or 6-page paper; the same article carries an ad for his paid launch agency (0x9.fun). The paper's 15-day claim can't pass its own n>=30 and 3-year gates as live proof. John's laoer research already predeclares and uses development/validation/holdout splits, and a six-LLM desk has no place in a fixed-rule bot.
original
i still don't understand why everyone is NOT building 24/7 trading agents with Grok Bot
it will probably make you a lot of money & definitely change your life...
i wrote a 6-page research paper on exactly how to find profitable strategies 24/7 with Grok Bot from scratch
simply paste this paper into grok bot, it will research the whole market & build you a 24/7 trading bot
this is the EXACT system I have been running for the past 15 days are so far results are INCREDIBLE:
PDF: https://drive.google.com/file/d/1q5bDURvU15flb41HP_CJW43LowNNQBjz/view (image-only; OCR below)
## Quoted post
Grok Bot builds MOST powerful trading agents right now
It gives you an army of agents that can do work for you 24/7
If you set it up correctly, you gain super powers
In this article, I cover how to build a team of 24/7 AI trading agents from scratch https://x.com/i/article/2117505964178067456
## Article: How to Use Grok Bot to Build a Team of AI Trading Agents
I will break down the EXACT framework to build a 24/7 AI trading team with Grok Bot that researches the market, finds strategies, backtests & trades on its own.
Let's get straight to it.
Bookmark This -
if you're building an AI/tech product, your product is not the problem, your distribution is. At 0x9.fun we make your AI product WIN on X with launches & long-form builds exactly like this one. DM me to launch your next campaign with 0x9 .
And if you’re a CREATOR, get paid to post JOIN HERE.
If you are building an AI trading system right now, DM me your setup and i will personally review the first 20. MOVE FAST.
Something big is happening on the timeline right now and MOST traders have not noticed it yet.
People are using Grok Bot to automate their entire lives: their inbox, their calendar, their research and their sales.
Founders are running whole companies with it. One person at the top, a team of Bots doing the real work below.
Each Bot gets its own computer in the cloud. It signs into the same tools you use and it never logs off.
It remembers your rules, works on a schedule, talks to the other Bots and only comes back when it needs your yes.
Now picture that same team pointed at the markets instead of an inbox.
An AI team that researches the market around the clock, spots the mispricing, turns it into a strategy, backtests it on years of data and builds the trading bot that executes it, all while you sleep.
That is not a chatbot anymore.
That is a full trading desk plugged straight into your TradingView, reading your charts, running your screener and watching your alerts 24/7.
Funds pay MILLIONS for exactly this structure: researchers, analysts, risk managers and execution traders, all on one floor, working in shifts.
You can build the same structure tonight, for the price of one monthly subscription, without writing a single line of code.
This article is the full build, from an empty app to a team that researches, analyzes and trades on its own.
By the end of this article you will know:
How to hire a complete AI trading team with one Chief of Staff running it
How to plug Grok Bot into your TradingView so it reads your charts, runs your screener and manages your alerts
How to make it build your own premium style TradingView indicators, without paying for a single one
How to wire in a broker and run the entire desk 24/7, first on paper, then with real money
What this setup CAN & CANNOT do, stated clearly enough to hold me to
Let's get into it.
Part 1: What Grok Bot Is and Why It Changes AI Trading
Every AI model or LLM before this worked the same way. You type. It answers. It stops.
It could explain a trade. It could never actually watch one.
Grok Bot works like a coworker and four things make the difference.
1. Its own computer:
Every Bot runs on a cloud computer. Your laptop can be closed, off or in another country. The Bot keeps working.
2. Routines:
You give a Bot a job once. It runs that job on a schedule: every morning, every 30 minutes, every Sunday. One Bot can own up to 50 routines.
3. Memory:
Bots remember your rules, your preferences and their past work. You never repeat yourself.
4. Teams:
Bots message each other, share context and coordinate in group chats. A Chief of Staff Bot manages the specialists and only escalates real decisions to you.
Now look at what a trading desk actually needs.
That last row matters more than people think. Grok Bot can search, read and monitor X.
X is where markets move first. CEOs post. Analysts post. Breaking news lands on X before it hits any headline. Your Bot can now sit on that feed for you, 24 hours a day.
Within days people were sharing ready-made Bot templates: research desks, sales teams, video editors and more.
It cost's $30 a month for a team that never sleeps. A single junior analyst costs more than that per hour.
Part 2: Hire Your AI Trading Team
Here is the MISTAKE almost everyone makes. They create one Bot and ask it to trade, "Watch BTC and tell me what to do."
That is not a system. That is a coin flip with extra steps.
Real desks split the work. One person finds ideas. One reads charts. One tests. One says no. One executes.
So we build exactly that.
Six Bots. One job each. You only talk to the top.
Every agent has one job. Every decision has to survive the entire route before it becomes a position.
That is the principle the whole desk is built on.
Step 1:
Download Grok Bot. Get it at x.ai/bot. It runs on macOS, Windows, Linux, iPhone and Android. Sign in with your X or Cursor account.
Step 2:
Write your Desk Rules. Before you hire anyone, write the rulebook.
Every Bot reads this before every decision. It is the one thing that keeps six agents acting like one desk.
Step 3:
Hire the Chief of Staff, Click the plus button, create a new Bot and paste this:
Step 4:
Give each specialist its job.
SCOUT finds what is moving and why.
CHART ANALYST turns stories into setups.
STRATEGIST writes the plan and proves it.
RISK OFFICER is the only Bot that can say no to everyone.
TRADER executes. Nothing else.
Step 5:
Lock down permissions. This step is not optional.
Bots are proactive by design. A prompt is a request, not a lock.
Open each plugin's tool list and switch off every action you do not need. Only the Trader should ever reach order tools. Everyone else reads.
Part 3: Connect TradingView, Build Indicators, Wire the Broker
A team without tools is just a group chat. Now we give it eyes, hands and a memory you can actually read.
Connection 1: TradingView, the eyes.
TradingView now has an official MCP server in public beta.
MCP is the open standard that lets an AI plug into real tools. Not talk about your charts. Actually READ them.
Through it, your Bots can:
pull quotes and price history
run the TradingView screener
read fundamentals, news, filings, earnings transcripts and the economic calendar
use indicators like RSI and moving averages
create and manage your watchlists and alerts, and read your alert log
You need a TradingView Essential plan or higher.
In any Grok Bot chat, type:
Use the words "custom server", or the Bot may search the plugin catalog instead.
The Bot shows you the URL. Check it and click Add it. A connect card appears. Click Authorize, sign in to TradingView and click Allow access.
Done. Type @ in any chat to attach TradingView to a task.
Test it straight away:
If a table comes back, your desk can see the market.
Connection 2: Your own premium style indicators.
This is the part people love (my personal favourite)
The TradingView connection reads indicators. It does not write Pine Script.
But your Bot has its own computer and its own browser.
So the Chart Analyst can write a custom indicator AND add it to your chart.
You approve the TradingView sign in once. The Bot does the rest.
Now you have an indicator built around YOUR rules, with a volatility filter most paid indicators do not even have.
People pay monthly subscriptions for indicators like this. Yours took one prompt.
Connection 3: The broker, the hands.
Alpaca has an official MCP server and a free paper trading account.
Create a free Alpaca account.
Open the paper trading dashboard.
Copy your paper API keys.
Then type:
The Bot runs the server on its own cloud computer & reports the tool count.
Test it:
Keep paper mode ON. Real money comes in Part 4.
Connection 4: The journal, the memory you can read.
You cannot read a Bot's memory directly. So make every Bot write everything down.
Add the Google Sheets or Notion plugin and ask the Chief of Staff to create a "Trade Journal" with these columns:
This sheet is your audit trail. When something goes wrong, you will know exactly which Bot said what, and why.
Connection 5: The 24/7 routines.
Now we turn the team into a desk that runs on its own.
Tell the Chief of Staff:
A day inside your desk.
Here is what this looks like when it runs.
You did not open a chart once. You read two messages and the desk did the rest.
Part 4: Paper to Real Money, Safely
Most AI trading bots die in the same place. The jump from paper to real money.
So we do not jump. We climb.
Moving to live.
Create live API keys in Alpaca. Ask the Bot to update the alpaca server with the new keys and set ALPACA_PAPER_TRADE to False.
Then add one rule to the Trader:
Grok Bot is built for this. It comes back to you only when something needs your approval.
The kill switch.
Tell the Risk Officer:
Then set the same limits inside your Alpaca account.
A prompt is a rule. A broker setting is a wall.
The weekly review.
Every Sunday the Chief of Staff compares paper or live results against the backtest.
Strategies drifting below expectations get paused. Failed ideas get logged so no Bot ever retests them by accident.
The desk gets sharper every single week. That is the real compounding.
The 5 mistakes that blow up AI trading desks:
The Honest Contract. Screenshot This.
This setup CAN:
research markets 24/7 with your laptop closed
monitor X live for every ticker on your watchlist
read TradingView quotes, screener, fundamentals, news, filings, watchlists and alerts
write custom indicators and add them to your TradingView charts
backtest ideas on its own cloud computer
place paper trades, and live trades with your approval, through Alpaca
run a full team with a Chief of Staff and log every single decision
This setup CANNOT:
trade at high frequency speed. Routines run minutes apart, not milliseconds
connect to TradingView without an Essential plan or higher
let you choose the model. Grok Bot routes models automatically
show you its memory. That is why the journal exists
run unlimited. Usage is a weekly pool, and Bot to Bot messages count
guarantee profit or turn a weak idea into an edge
That second list is not hedging. It is the reason the first list is believable.
Closing
For years we asked AI to trade from a chat box. It was never the right shape.
A trading desk is not one genius. It is a team, a process and a schedule that never breaks.
Grok Bot finally gives you all three.
One Chief of Staff. Five specialists. TradingView for eyes. Alpaca for hands. A journal for memory. Routines that never sleep.
Build it once on paper. Let it prove itself. Then let it work.
In my previous article i broke down how to build a millisecond speed 24/7 HFT system with Jev. This one gives you team that runs everything around it.
So here is the question.
Is your AI still waiting for you to open a chat window or is it already running a desk while you sleep?
There is no wrong answer. But there are very revealing ones.
## Paper (OCR; equations and tables are rough)
### Page 1
How to Build a One-Person Hedge Fund With Grok Bot
That Autonomously Discovers Profitable Strategies 24/7
A Field Report on Running Research, Backtesting, Risk, and Execution
Through a Team of Always-On Agents With Their Own Cloud Computers
Abstract— A one-person trading operation has always
been limited by one resource: the operator’s attention.
Markets move around the clock, news breaks at any hour,
and every serious strategy demands research, testing, risk
review, and execution that no single person can sustain.
Chat-based AI did not solve this, because a chat model
answers once, forgets, and stops when the window closes.
Grok Bot changes the unit of work from a conversation to
an employee. Each Bot runs on its own cloud computer,
keeps persistent memory, executes scheduled routines, and
coordinates with other Bots in a shared team. We document
a six-agent desk built on this primitive: a Chief of Staff
that orchestrates, a Scout that monitors news and live X
activity, a Chart Analyst connected to TradingView through
its official Model Context Protocol server, a Strategist that
converts setups into tested rules, a Risk Officer with absolute
veto power, and a Trader that executes only approved orders
through a broker connection, beginning in paper mode. We
define a strategy as profitable only when it remains positive
after costs on data it never saw, survives a deflated Sharpe
test, and holds across walk-forward windows. The central
claim is that a hedge fund is not a building full of people.
It is a process, and that process can now run autonomously
under one operator.
Index Terms— Grok Bot, autonomous agents, multi-agent
orchestration, one-person hedge fund, Model Context
Protocol, TradingView, Pine Script, walk-forward validation,
deflated Sharpe ratio, fractional Kelly, paper trading, risk
veto, scheduled routines.
I. INTRODUCTION: THE ATTENTION CEILING
Every independent trader eventually hits the same wall.
The market never closes, but the trader does. Opportunities
appear before the open, during lunch, and after midnight,
and each one requires the same chain of work: find it, read
the chart, test the rule, check the risk, place the order. A
single person can run that chain a few times a day. A desk
runs it continuously.
The usual answer has been to automate one link of the
chain, a screener here, an alert there, a script that places
orders. Each tool helps, but none of them thinks, and none
of them coordinates with the others. The operator remains
the glue, and the operator’s attention remains the ceiling
on everything the system can do.
This paper removes the operator from the loop of
routine work and keeps the operator at the top of the
loop of decisions. The result is a desk that researches,
tests, and trades around the clock, and asks for a human
only when capital is at stake.
A. Why a Chatbot Cannot Run a Desk
A chat model asked to trade receives a question, produces
a confident answer, and stops. It has no schedule, so it cannot
watch the market. It has no persistent memory, so it repeats
last week’s rejected idea as if it were new. It has no tools, so
it describes a chart it has never opened. It has no colleagues,
so the same model that proposes a trade also approves it.
Each limitation is fatal on its own. Together they explain
why chat-based trading has remained a demonstration
rather than a desk.
B. The Six-Agent Desk Definition
This paper defines the desk as six agents with one job
each. The Chief of Staff enforces process and is the only
agent that speaks to the operator. The Scout finds what is
moving and why. The Chart Analyst turns stories into
scored setups. The Strategist writes exact rules and proves
them on history. The Risk Officer can reject any plan. The
Trader executes only plans stamped as approved. No agent
moves real capital without operator approval.
C. One Chatbot Versus the Six-Agent Desk
Table I contrasts a single chatbot asked to trade with the
same intelligence divided across the six-agent desk, across
the six functions that decide whether a strategy survives a
live market. The difference is not intelligence. It is whether
every function has an owner, a schedule, and a record.
TABLE I
ONE CHATBOT VS THE SIX-AGENT DESK
Function One Chatbot Six-Agent Desk
Market watch Only when asked — Scout, every 30 min
Chart reading § From memory TradingView MCP
Strategy testing Untested Strategist backtests
Risk control Self-approved Risk Officer veto
Execution Manual Trader, paper first
Memory Resets each chat _— Persistent + journal
D. What This Paper Delivers
Section II specifies the desk architecture, the agent
routines, and the operating rules every agent reads. Section
III presents the strategy discovery loop and its statistical
gates. Section IV details the TradingView and broker
integration, including the custom indicator mathematics.
Section V covers position sizing, risk limits, and the kill
switch. Section VI covers the paper-to-live promotion
ladder, daily operation, and references.
### Page 2
II. DESK ARCHITECTURE
A. The Primitive: An Agent With a Computer
The desk is built on one primitive. A Grok Bot is an
agent with its own cloud computer, persistent memory,
and scheduled routines. It signs into tools the way a
person does, runs whether or not the operator’s laptop is
open, and can own up to fifty routines. Bots message
each other, share context, and work inside group chats.
The architecture treats each Bot as an employee with a
single job description, not as a general assistant.
B. The Decision Path
Let x; denote the market state observed at time ¢,
including prices, volume, news, and live X activity. Every
action the desk takes is the output of a fixed composition
of agents, applied in order:
a, = T(R(S(C(Se(x:)) )) (1)
where Sc is the Scout, C the Chart Analyst, S the
Strategist, R the Risk Officer, and T the Trader. No agent
may skip a stage, and every stage may return a null action,
in which case the path terminates and the reason is logged.
The composition matters because each stage is an
independent filter. If f; is the probability that stage i
passes a bad idea, the probability that a bad idea reaches
execution is:
k
P(bad trade executed) = Ils (eh 2nk @
i=1
With four filtering stages each passing a bad idea thirty
percent of the time, the desk executes a bad idea with
probability 0.34 = 0.0081, fewer than one in a hundred. A
single agent that both proposes and approves offers no
such protection.
TRADE JOURNAL
Fig. 1. Six-agent desk architecture and decision path.
C. Desk Rules as Shared State
Every agent reads one document before every action:
the Desk Rules. It fixes the tradable markets, capital,
maximum risk per trade, maximum daily loss, maximum
open positions, holding period, and the instruments the
desk must never trade. It also fixes two behavioural rules:
every claim carries a source and a timestamp, and no
agent may invent a price, a number, or a result. The Desk
Rules act as shared state, so six agents behave as one
desk rather than six opinions.
D. Routines and the 24-Hour Schedule
Autonomy comes from routines. Each agent owns
scheduled jobs that run on its own computer, so the desk
operates continuously without operator prompts. Table II
lists the schedule used in this study, expressed in US
Eastern Time.
TABLE I
AGENT ROUTINE SCHEDULE (ET)
Agent Schedule Output
Scout 06:30, then every 30 min Ranked findings
Chart Analyst 08:30, then hourly Setups scored 7+
Strategist 21:00 nightly Tested trade plans
Risk Officer Every order, 16:15 Approve or veto
Trader Every 15 min in session Orders and fills
Chief of Staff 07:30, 16:30, Sun 18:00 Briefs and review
E. The Handoff Protocol
Agents communicate in one fixed format: TICKER,
IDEA, EVIDENCE, NEXT STEP. A fixed format removes
ambiguity between agents and makes every message
machine-checkable. An idea that arrives without evidence
is returned to its sender. An idea that arrives at the Trader
without an APPROVED stamp from the Risk Officer is
refused.
F. Memory and the Trade Journal
Grok Bot memory persists across sessions but cannot
be inspected directly. The desk therefore writes every
idea, score, plan, risk decision, order, fill, and lesson to an
external Trade Journal. The journal is the audit trail, the
source of calibration data, and the record that prevents
any rejected idea from being retested by accident.
### Page 3
IIL THE STRATEGY DISCOVERY LOOP
A. From Signal to Hypothesis
Discovery begins with the Scout, which records
every market event with a source and a timestamp.
Each night the Strategist converts the day’s findings
into at least ten testable hypotheses. A hypothesis is
admissible only if it names its edge, the reason it
should work, and the counterparty who loses when it
does. Ideas without a stated edge are discarded before
any computation is spent on them.
B. The Setup Score
The Chart Analyst scores every candidate on five
components read from TradingView: trend alignment,
proximity to a key level, momentum, relative volume,
and catalyst strength. Each component s; is scored
from 0 to 10 and combined with fixed weights w;:
5 5
S=)owjs;, }iwjs=1, 5; €[0,10] (3)
i=l i=l
Only setups with S > 7 advance to the Strategist.
Fixing the weights in the Desk Rules prevents the score
from drifting toward whatever the agent finds persuasive
on a given day.
C. Backtest Statistics
Every plan is backtested on at least three years of
history. Gross results are meaningless without costs, so
each of the n round trips is charged a cost c,,; covering
commission, spread, and slippage:
Rnet = R gross —N-Cry (4)
Risk-adjusted performance is measured by the annualised
Sharpe ratio of daily net returns r, and downside by the
maximum drawdown of the equity curve V;:
SR = —. 252 (5)
Oo,
P,-V,
MDD = max — , P,=maxV, (6)
t t SSI
D. Walk-Forward Validation
A single backtest describes one path through history.
The desk instead splits the data into rolling windows, fits
rules on each training window, and evaluates them
untouched on the following test window. Rules are never
adjusted after the test window is seen. A strategy that only
works in one window is classified as luck, not edge.
E. The Multiple-Testing Problem
An agent that tests ten ideas a night tests over three
thousand a year, and some will look excellent by chance
alone. Among N strategies with no true edge, the best
observed Sharpe ratio grows approximately as:
E[ max SR] * oOsp- V2InN (7)
The desk therefore evaluates every survivor with the
deflated Sharpe ratio, which corrects for the number of
trials, the sample length T, and the skewness y; and
kurtosis y4 of returns:
(SR — SRy) VT —1
DSR = 0} ———————— }_ (8)
=
1—ysn+ 4 spe
where SRo is the expected maximum Sharpe under N
trials from (7) and ® is the standard normal distribution
function.
EF. The Gates
A strategy is called profitable in this paper only if it
passes every gate in Table III. Failing any single gate
sends it to the graveyard.
TABLE Ill
STATISTICAL GATES FOR A PROFITABLE STRATEGY
Gate Threshold Purpose
Trade count n>30 Statistical meaning
Out-of-sample Rnet > 0 Profit after costs
Deflated Sharpe DSR > 0.95 Luck correction
Walk-forward > 70% windows positive Stability
Latest year Positive Current regime
Parameter plateau Neighbours within 20% No lucky setting
Noise test 1-bar delay, 2x slippage Real-world friction
### Page 4
IV. TRADINGVIEW AND BROKER
INTEGRATION
A. MCP as the Interface
Agents act on markets through the Model Context
Protocol, an open standard that lets an AI system call
external tools. TradingView’s official MCP server, in
public beta, exposes quotes, price history, the screener,
fundamentals, news, filings, the economic calendar,
standard indicators, watchlists, alerts, and the alert log.
It requires an Essential plan or higher, serves delayed
data during the beta, and does not author Pine Script.
The desk treats it as the eyes of the system, not as an
execution venue.
B. Connection Procedure
The connection is created in chat. The operator
instructs a Bot to add a custom MCP server named
tradingview at the TradingView MCP endpoint, reviews
the address, authorises the connection, and grants access
from the TradingView account. Once connected, any agent
can attach TradingView to a task, and the Chart Analyst
reads every chart through it.
C. The Desk Signal Indicator
Because each Bot controls its own browser, the Chart
Analyst can author a custom indicator, paste it into the
Pine Editor, and attach it to the operator’s charts. The
indicator used in this study, Desk Signal, combines four
components. Trend is measured by an exponential moving
average ribbon with lengths n € {8, 21, 50, 200}:
2
=— 9
= n+1 0)
EMA, = a@P, +(1—a@)EMA,_, ,
The trend is bullish only when the ribbon is fully stacked,
EMAg > EMA}, > EMAsy > EMApq9, and bearish under
the reverse order. Momentum is measured by the 14-period
relative strength index:
100 G
, RS,=~*
1+RS, Dm
RSI, = 100 — (10)
where G4 and L}4 are the average gain and average loss
over fourteen bars.
A buy signal requires a breakout above the prior twenty-
bar high on expanded volume, confirmed by momentum
and a bullish trend:
BUY, = 1[C, > max H,_29:4-1 AV; = 1.5V29 A RSI, > 50]
e8))
Signals are suppressed in disorderly markets. With true
range TR, = max(H, — L,, |H, — C,-4|, |L, — C;-1]), the
filter blocks every signal when:
ATR 44,4 > Qo.90 (ATRi4 over the last 100 bars) (12)
Listing 1 shows the core of the indicator as written by
the Chart Analyst.
//@version=6
indicator("Desk Signal", overlay=true)
e8 = ta.ema(close, 8)
e21 = ta.ema(close, 21)
e5@ 0= ta.ema(close, 50)
e200 = ta.ema(close, 200)
up = e8 > e21 and e21 > e50 and e50 > e200
dn = e8 < e21 and e21 < e50 and e50 < e200
vol0K = volume >= 1.5 * ta.sma(volume, 20)
atr = ta.atr(14)
calm = atr <= ta.percentile_linear_interpolation(atr, 100, 90)
rsi = ta.rsi(close, 14)
buy = up and close > ta.highest(high, 26)(1] and volOK and
rsi > 50 and calm
sell = dn and close < ta.lowest(low, 26)[(1] and volOK and
rsi < 50 and calm
plotshape(buy, "BUY", shape.labelup, location.belowbar, text="BUY")
plotshape(sell, “SELL”, shape. labeldown, location.abovebar, text="SELL")
alertcondition(buy, “Desk BUY", “Desk Signal BUY")
alertcondition(sell, “Desk SELL", “Desk Signal SELL")
Listing 1. Desk Signal core logic in Pine Script v6.
D. Broker Connection and Tool Isolation
Execution runs through Alpaca’s official MCP server,
launched on the Bot’s own computer in paper mode. Order
tools are the most dangerous capability in the system, so
access is isolated by role, as shown in Table IV. Only the
Trader can reach order tools.
TABLE IV
TOOL ACCESS BY AGENT
Agent TradingView Broker
Scout News, calendar None
Chart Analyst Charts, alerts None
Strategist Price history None
Risk Officer Read only Account read
Trader Alert log Orders (paper)
Chief of Staff Read only Account read
### Page 5
V. POSITION SIZING AND RISK
A. Fixed-Fractional Sizing
A strategy that passes every gate can still destroy an
account if it is sized badly. The desk sizes every position
so that the loss at the stop equals a fixed fraction f of
current equity E. For an entry price P, and stop price P,,
the position quantity is:
f-E
p= eS, (13)
|P. — P,|
With f = 1% and E = $10,000, a trade with a $2 stop
distance buys 50 shares, and the worst planned loss is
$100 regardless of the price of the instrument.
B. Volatility Scaling
Stops are placed at a multiple k of the average true
range, so stop distance widens automatically when
markets become volatile. Substituting into (13):
f-E
a=
= 14
k- ATRy, a")
Position size therefore shrinks as volatility rises, without
any discretionary judgment from the agents.
C. Fractional Kelly as an Upper Bound
For a strategy with out-of-sample win rate p and
payoff ratio b, the average win divided by the average
loss, the Kelly fraction is:
Lop
fo =p ;
(15)
Full Kelly is too aggressive for estimated inputs, so the
desk uses a quarter of it, capped by the Desk Rules:
p=min( ©, far)
For p = 0.55 and b = 1.5, f* = 0.25 and quarter Kelly is
6.25%, but the 1% cap binds. Kelly acts as a ceiling,
never as a target, and its inputs come only from
out-of-sample trades.
(16)
D. Correlation and Portfolio Heat
Two positions in correlated instruments are one bet
sized twice. With position weights w and return
covariance matrix Z, portfolio volatility is:
(17)
The Risk Officer treats any two positions with correlation
above 0.7 as a single bet, and limits total open risk, the
portfolio heat H, across all n positions:
H=)-f; < Hmax = 5% (18)
Ms
~
ll
=
E. Event Risk
Scheduled events produce gaps that no stop can protect
against. The Risk Officer blocks new positions before
earnings, central bank decisions, and inflation releases, and
flags every open position whose target date crosses one.
F. The Kill Switch
Some failures require the desk to stop entirely. Table V
lists the triggers. When any trigger fires, the Risk Officer
cancels open orders, closes positions where specified,
pauses the Trader, and alerts the operator.
TABLE V
KILL SWITCH TRIGGERS
Trigger Threshold Action
Daily loss > 2% of equity Flatten, pause desk
Losing streak 3 ina row Pause Trader
Stale data No update in 2 bars Cancel all orders
Slippage Fill > 0.5% from plan Pause, review
Rule breach — Any agent off Desk Rules Freeze desk
Strategy drift Below Monte Carlo Sth pct 9Demote strategy
G. Layered Enforcement
Every limit is enforced in three independent layers: the
Desk Rules read by each agent, tool isolation that removes
order access from every agent except the Trader, and hard
limits configured inside the broker account itself. A prompt
is a rule. A broker setting is a wall.
### Page 6
VI. FROM PAPER TO LIVE CAPITAL
A. Fixed.Praction Ladder
Most automated trading systems fail at one transition:
the move from simulated to real capital. The desk never
makes that move in a single step. Every strategy climbs the
five rungs in Table VI, and a failure at any rung returns it
to the rung below.
TABLE VI
THE PAPER-TO-LIVE PROMOTION LADDER
Rung Condition to Advance Capital
1. Paper 2 weeks, all trades logged Simulated
2. Match Paper consistent with backtest Simulated
3. Live tiny Clean fills, no breaches 10% of plan
4. Approval gate Operator approves each order Planned size
5. Full rules Weeks of clean journal Planned size
B. The Consistency Test
Rung two asks whether paper results are consistent
with the backtest. With n paper trades of mean return 7,,
and a backtest mean 7, with per-trade standard deviation
o,, the desk computes:
r=
~ on /J/n
Zz (19)
A strategy advances only when n > 20 and z>-2.A
strongly negative z means live conditions are destroying
the edge the backtest promised, usually through costs,
latency, or a changed regime, and the strategy is returned
to the Strategist.
C. Daily Operation
The operator’s working day reduces to two messages
and a set of approvals. The morning brief summarises
the market, ranks, and surviving setups, and lists every
veto with its reason. Live orders arrive as a fixed card
of ticker, side, entry, stop, target, size, risk, and rationale,
and silence is treated as rejection. Each Sunday the Chief
of Staff compares live results with backtests, demotes
decaying strategies, and archives failed ideas in a graveyard
so they are never retested.
D. Limitations
The desk is not a high-frequency system. Routines run
minutes apart, so edges that decay within seconds are out
of reach. TradingView data is delayed during the MCP
beta. The operator cannot select the underlying model,
which Grok Bot routes automatically, and cannot inspect
agent memory directly, which is why the external journal
is mandatory. Usage draws on a weekly pool shared by
all agents. Finally, the statistical gates reduce the probability
of trading luck, but they cannot guarantee future profit, and
no architecture can turn a strategy without edge into one
with it.
VII. CONCLUSION
A hedge fund has never been defined by its building or
its headcount. It is defined by a process: research, testing,
risk control, and disciplined execution, running without
interruption. This paper shows that the process can now
be assigned to a team of always-on agents, each with one
job, one schedule, and one record, while a single operator
keeps authority over capital. The edge still has to be found.
The difference is that the search no longer stops when the
operator does.
REFERENCES
[1] xAI, “Introducing Grok Bot,” x.ai, 2026.
[2] TradingView, “TradingView MCP server public beta,”
TradingView Blog, 2026.
[3] Alpaca Markets, “Alpaca MCP server,” GitHub repository,
2026.
[4] Model Context Protocol, “Specification,” modelcontext-
protocol.io, 2025.
[5] D.H. Bailey and M. Lépez de Prado, “The deflated Sharpe
ratio: Correcting for selection bias, backtest overfitting, and
non-normality,” J. Portfolio Management, vol. 40, no. 5,
pp. 94-107, 2014.
[6] D.H. Bailey, J. M. Borwein, M. Lopez de Prado, and Q. J. Zhu,
“Pseudo-mathematics and financial charlatanism: The effects
of backtest overfitting on out-of-sample performance,” Notices
of the AMS, vol. 61, no. 5, pp. 458-471, 2014.
[7] J. L. Kelly, Jr., “A new interpretation of information rate,”
Bell System Technical Journal, vol. 35, no. 4, pp. 917-926,
1956.
[8] W. F. Sharpe, “The Sharpe ratio,” J. Portfolio Management,
vol. 21, no. 1, pp. 49-58, 1994.
[9] J. W. Wilder, New Concepts in Technical Trading Systems.
Greensboro, NC, USA: Trend Research, 1978.
[10] R. Pardo, The Evaluation and Optimization of Trading
Strategies, 2nd ed. Hoboken, NJ, USA: Wiley, 2008.
[11] TradingView, “Pine Script v6 reference manual,” Trading View,
2026.