In last week's post about why AI trading bots are a scam, we went through exactly why asking an LLM to predict price direction is a broken idea. LLMs are next-token prediction engines trained on text. They cannot tell you where BTC is going tomorrow. They have no live market data, no understanding of order flow, and no measurable expectancy.
That post covered what not to do. This one covers what to actually do.
LLMs are genuinely useful for three things in trading: explaining concepts without dumbing them down, organizing data without making arithmetic errors, and analyzing patterns you point them at. None of those things involve predicting price. All of them are legitimately valuable for building a real trading system.
Here is the exact four-step process we use and teach.
Step 1: Understand what a trading system actually is
Before you can build or test a system, you need a precise definition. A trading system has exactly four components. Not three, not five. Four.
Entry condition
The specific, objective rule that triggers a trade. No ambiguity. You can write it as code.
Stop loss
The price level where you exit if the trade is wrong. Set before entry, never moved against you.
Take profit
The price level where you exit if the trade is right. Or a trailing rule. Either way, defined in advance.
Position size
How many units you trade, calculated from the distance to your stop and the dollar amount you are willing to risk.
Use this prompt to get a concrete walkthrough from an LLM:
Prompt: Explain a trading system with exact math
You are an experienced algorithmic trader. Explain what a complete trading system is using a specific moving average crossover example on BTC/USDT. Include:
- The exact entry condition using a 20 EMA and 50 EMA crossover on the 4-hour chart
- How to set a stop loss below the previous swing low
- A 2R take profit target
- The exact position sizing math for:
- Account size: $10,000
- Maximum risk per trade: 1% ($100)
- BTC price at entry: $65,000
- Stop loss distance: $1,200
Show all arithmetic step by step. State the number of BTC to buy and the dollar value of the position.
When you run that prompt, you will get an output something like this: risk amount is $100, stop distance is $1,200, so position size is 100 / 1200 = 0.0833 BTC, which at $65,000 is a $5,416 position. That is the math a real trader does before every trade. The LLM does it accurately and explains the reasoning. That is the correct use of the tool.
Step 2: Pick a simple system and log every trade
The most common mistake new systematic traders make is skipping the logging step. They paper trade for two weeks, feel like the system is working, and go live. Then they have a 10-trade losing streak and have no idea whether it is bad luck or a broken system, because they have no data.
You need a minimum of 100 trades logged with identical rules before the numbers mean anything. Below 100, the win rate and expectancy figures could easily be statistical noise.
Your trade journal needs to capture every variable that matters. Use this prompt to set one up:
Prompt: Set up a trade journal CSV
Create a CSV template for a systematic trading journal. I need to track every paper trade for a moving average crossover strategy on BTC/USDT. Include these columns:
trade_number, date_time_entry, instrument, direction (long/short), entry_price, stop_price, take_profit_price, position_size_units, position_size_dollars, risk_amount_dollars, entry_reason, date_time_exit, exit_price, exit_reason, result (win/loss/breakeven), pnl_dollars, pnl_r, notes
After the template, explain what R-multiple means and why I should track pnl_r instead of only pnl_dollars.
The R-multiple column is the most important one. R normalizes every trade by the amount risked. If you risk $100 and make $250, that is a 2.5R win. If you risk $100 and lose $100, that is a 1R loss. R-multiples let you compare a $500 trade and a $5,000 trade on equal footing, which is essential once you start varying position sizes.
Paper trade the system until you have at least 100 trades. Follow the rules exactly. Do not override the entry. Do not move the stop. The point is not to see if you can make the system work by feel -- the point is to let the rules run and see what the data says.
Step 3: Calculate your expectancy
After 100 trades, you have real data. Now you calculate expectancy. This is the single most important number a systematic trader produces. It tells you how much you expect to make per dollar risked, on average, across all future trades -- assuming the market conditions that produced the sample continue.
Expectancy formula and example calculation
Expectancy = (Win Rate x Avg Win in R) - (Loss Rate x Avg Loss in R)
Example: Win rate: 45% Average win: 2.1R Loss rate: 55% Average loss: 1.0R
Expectancy = (0.45 x 2.1) - (0.55 x 1.0) = 0.945 - 0.55 = 0.395R per trade
A positive expectancy means the system makes money over time, assuming you follow the rules consistently. A 0.395R expectancy means for every $100 you risk, you expect to net $39.50 on average across many trades. The number sounds small. Across 500 trades with 1% risk on a $20,000 account, that is roughly $39,500 in expected profit.
Use this prompt to run the calculation on your actual journal data:
Prompt: Calculate expectancy from trade data
I have a trading journal with 120 trades. Here is a summary of my data:
[paste your actual numbers here]
Calculate:
- Win rate and loss rate
- Average win in R and average loss in R
- Expectancy in R per trade
- Total R earned over the sample
- Maximum consecutive losses
- Largest single loss in R and largest single win in R
Then tell me whether this expectancy is meaningful given the sample size, and what the minimum sample size should be before drawing conclusions.
If your expectancy is negative after 100 trades, that does not necessarily mean the system is broken -- it might mean the market regime during your sample was unfavorable, or you had execution errors, or the sample is too small. What it does mean is that you do not go live yet.
Step 4: Analyze results and iterate
Expectancy is the summary number. Before you iterate on the system, you need to understand where the edge comes from and where it breaks down. A positive overall expectancy can mask a system that works well in certain conditions and fails badly in others.
Prompt: Break down performance by conditions
I have a trade journal with 150 completed trades in CSV format. Please analyze:
- Win rate broken down by day of week (Monday through Friday)
- Win rate broken down by trading session (Asian 00:00-08:00 UTC, London 08:00-13:00 UTC, New York 13:00-21:00 UTC)
- The 5 largest winning trades -- what do they have in common in terms of entry reason and session?
- The 5 largest losing trades -- what do they have in common?
- Maximum consecutive losing streak and the drawdown in R during that streak
- If I removed all trades taken during [specific session], what would expectancy become?
[paste CSV data here]
Based on this analysis, suggest two specific rule changes I could test to improve expectancy, and explain the reasoning for each.
This is the iteration loop: analyze the data, identify the weakest part of the system (usually a specific session or market condition where the edge disappears), test a filter that removes those conditions, and re-run the backtest. Each iteration either improves the system or teaches you something about the market.
Common findings from this analysis:
- Trend-following crossover systems often perform poorly in the Asian session (low volume, choppy price action) and well in the London and New York sessions
- Wednesday and Thursday tend to have better trend persistence than Monday and Friday
- The largest losing trades often share a specific pattern (entering into news events, taking signals at major resistance, low ATR environments)
The LLM is useful here because it can run the segmentation quickly without writing pandas code yourself, and it can propose conditional filters in plain English that you can then test.
Why this process works
The LLM is not predicting price in any of these steps. It is doing four specific things it is genuinely good at:
Explaining concepts clearly
LLMs are exceptional teachers. They can explain position sizing, expectancy, R-multiples, and drawdown in plain language with concrete numbers, at whatever level of detail you need. This is reading and language comprehension -- exactly what they are trained for.
Organizing data structures
The journal template prompt produces a complete, well-structured CSV schema in seconds. Building that from scratch means either knowing what you need in advance (which requires experience) or forgetting columns and having to rebuild later.
Doing math without errors
Expectancy, position sizing, drawdown in R, consecutive loss probability -- all of this is arithmetic that humans do slowly and occasionally get wrong. LLMs do it accurately and show the steps. This alone eliminates a meaningful category of mistakes.
Pattern recognition in your data
Note the distinction: pattern recognition in your trade data, not in the market. Once you give the LLM a structured dataset of your own trades, it can find segments and correlations that would take you hours to identify manually. It is analyzing records you already have -- not predicting what price will do next.
None of this requires the LLM to have market knowledge, real-time data, or predictive ability. It requires the LLM to be competent at reading, arithmetic, and data summarization. Those are genuine strengths.
The traders who consistently make money are not the ones with the best predictions. They are the ones with a tested system, a measured expectancy, proper risk management, and the discipline to follow rules when it is uncomfortable. An LLM can help you build that system faster. It cannot replace the discipline required to run it.
Summary checklist
- Use an LLM to understand the four components of a trading system: entry, stop loss, take profit, position size
- Ask for concrete math examples -- position sizing, risk calculations -- not price predictions
- Set up a complete trade journal CSV before taking a single paper trade
- Log a minimum of 100 trades with identical rules before analyzing results
- Calculate expectancy: (Win Rate x Avg Win in R) minus (Loss Rate x Avg Loss in R)
- Break down performance by session and day of week to find where the edge lives and where it disappears
- Use the LLM to analyze your data, not to predict the market
- Iterate on rules based on data, test the change, and re-measure expectancy
- Do not go live until expectancy is positive across at least 100 trades with consistent execution
Ready to automate a proven system?
Once you have paper traded a system through 200+ trades and confirmed a positive expectancy, the next step is automation. We take strategies with proven edge and turn them into production trading software with proper risk controls, monitoring, and execution infrastructure. Book a free 30-minute diagnostic to discuss your setup.
Frequently asked questions
Can ChatGPT predict stock or crypto prices?
No. LLMs like ChatGPT and Claude have no real-time market data, no execution capability, and no special insight into order flow. They cannot predict where price goes next. They are useful as research assistants for building trading systems, organizing trade data, and doing math without errors.
What is trading expectancy and how do I calculate it?
Expectancy tells you how much you expect to make per dollar risked, on average, over many trades. The formula is: (Win Rate x Average Win in R) minus (Loss Rate x Average Loss in R). A positive number means the system makes money over time. You need at least 100 trades to calculate a meaningful expectancy.
How many trades do I need before I know if a trading system works?
A minimum of 100 trades following the exact same rules. This gives you enough data to calculate expectancy, win rate, average win, average loss, and maximum drawdown with reasonable statistical significance. Fewer than 100 trades and the results could easily be random.
What are R-multiples in trading?
R-multiples normalize every trade by the amount risked. If you risk $100 on a trade and make $250, that is a 2.5R win. If you risk $100 and lose $100, that is a -1R loss. R-multiples let you compare trades fairly regardless of position size, making them more useful than dollar P&L for evaluating system performance.
Ready to automate a proven system?
Once you have paper traded a system through 200+ trades and confirmed a positive expectancy, the next step is automation. We take strategies with proven edge and turn them into production trading software with proper risk controls, monitoring, and execution infrastructure. Book a free 30-minute diagnostic to discuss your setup.