A paper trading bot runs your strategy on live market prices with simulated money, and records every order as if it had been filled. To paper trade crypto with AI, you describe the strategy to an agent such as Claude or ChatGPT, it writes and backtests the code, and a platform runs that code on paper while you watch. Paper trading shows whether the strategy behaves live the way it did in the backtest. It does not show that it will make money.
What paper trading is
A backtest replays the past. Paper trading runs forward: the bot decides on bars that did not exist when you wrote the rules, so nothing can be tuned to them. That is why it is also called a forward test. No exchange account is involved and no money can be lost.
It sits between two other stages. Before it comes the backtest, which is cheap and fast but easy to fool. After it comes live trading, which is the only real test and the only one that costs money. Paper is where you find the problems that are free to find.
What paper trading proves
- The code runs unattended. It does not crash at 3 a.m., and it handles a gap, a quiet weekend and a violent hour.
- Signals arrive when the backtest says they should. If the live bot enters on bars where a backtest of the same period does not, something is wrong with the code or the data.
- The trade frequency is what you expected. A strategy that was meant to trade twice a week and trades twice a day is a different strategy.
- Fills at live prices look like the simulated ones, at least for small market orders.
- You get new out-of-sample data. Every paper week is a week the strategy was never tuned on.
What it does not prove
Paper results flatter you in three ways, and it helps to name them before you read any number.
- Fills. A paper order never reaches the real order book. It does not move the price, does not wait in a queue and is never partly filled or rejected. For a small order on a liquid pair the difference is small. For a large order, a thin coin or a limit order that needs to be reached first, it can be most of the edge.
- Slippage. A good simulator fills a market order at a realistic price instead of the last trade, but it still cannot know what happens in the moments around a real order during a fast move. Strategies with many small trades are hurt most, because the gap is paid on every entry and exit.
- Emotions. Watching a simulated account fall 15% costs nothing. Watching your own money do it is different, and most people who stop a bot do so at the worst point of a drawdown. Paper cannot tell you whether you will leave the bot alone.
It also does not prove profitability. A few good weeks on paper are a small sample, and one market regime. Trading is risky, and neither a backtest nor a paper record guarantees a future result.
How long to run it
Count trades, not days. A useful paper run has dozens of trades and at least one drawdown, so you have seen the strategy lose and recover. How long that takes depends on how often it trades.
Here is a real case. An EMA cross trend strategy (long when EMA 20 crosses above EMA 80, with a 3 ATR trailing stop) on BTCUSDT perpetual, 4h bars, from 2022-01-01 to 2026-01-01 came back from OptAlgo with honesty grade A: return +70.8% after fees and funding, max drawdown -15.0%, Sharpe 0.85, 55 trades, win rate 40%. That is one backtest, in-sample, so it is a description of the past and not proof of an edge.
Look at the trade count. 55 trades in four years is about one trade a month. Three weeks on paper would show one trade, maybe none. A win rate of 40% also means that five losses in a row are ordinary: for any five trades the chance is about 8%. So a short paper run of a slow strategy tells you the plumbing works and almost nothing about the edge. A strategy on 15-minute bars can reach dozens of trades in a couple of weeks; a 4h trend strategy needs months.
If you are not willing to wait that long, use the paper stage for what it can show in the time you have (does the code run, do the entries match) and lean on out-of-sample tests for the rest. A walk-forward test covers years of unseen data in one run.
How to compare paper results with the backtest
Do not compare your paper month with the original backtest's headline return. The periods are different, so the numbers will differ even if everything works. Compare like with like, in this order.
- Same period, trade by trade. Backtest the same code over exactly the days the bot ran, then line the two trade lists up. Entries and exits should fall on the same bars. A trade that exists on one side only is a bug to find, not noise.
- Fill prices. For the trades that match, look at the difference between the paper fill and the backtest fill. A small, random difference is fine. A difference that is always against you is slippage the backtest does not model.
- Behaviour against the long backtest. Trades per week, average holding time, win rate and average trade should sit inside the range the long backtest showed. With few trades these will wobble, so look for large gaps only.
- Drawdown. A paper drawdown that is already deeper than the worst one in a multi-year backtest is a warning. A drawdown inside that range is the strategy doing what it said it would.
On OptAlgo the first two steps are one feature. Compare with backtest runs a backtest of the same code over the period the bot ran and matches it against the real paper trades, trade by trade. Your agent can call it and report the trades that differ.
How to start a paper bot with an AI agent
You need a free OptAlgo account (sign up here) and an AI agent. No exchange account and no card are needed for paper.
Step 1: connect the agent. In Claude.ai or ChatGPT, add https://mcp.optalgo.com/mcp as a custom connector, sign in and approve. In Claude Code, add it from the terminal, then type /mcp and sign in:
claude mcp add --transport http optalgo https://mcp.optalgo.com/mcp
The connector uses OAuth, so no key is pasted into the chat. It is paper-only by design: it can backtest, deploy on paper and open and close paper positions, and it can never place a live order or change a live bot. The security page lists what each kind of access can do.
Step 2: describe the strategy and backtest it. Say the idea in one sentence, for example “go long BTC when the 20 EMA crosses above the 80 EMA on 4h and trail a 3 ATR stop”. The agent writes the code, validates it for free and runs the backtest. Ask it to report the honesty grade (A to D) and the overfit verdict before any return.
Step 3: deploy on paper. A backtest graded A to C can be deployed. The deploy gate always refuses a losing backtest, and refuses one judged likely overfit unless an out-of-sample research run of the same code passed. OptAlgo then runs your code at every bar close in a sandbox, on paper, as a bot in My Bots marked Runs on OptAlgo. Paper fills use the live order book, so a market order gets a realistic price and not just the last trade.
Step 4: watch it. Ask the agent things like “list the paper bot's trades this week” or “compare the paper bot with its backtest and show me the trades that differ”. You can also read every decision in My Bots, pause the bot or download the exact code it runs.
The Free plan includes one hosted strategy on paper and 200 paper trades a month. Paper trading does not use backtest quota; only running a backtest does. Limits for the other plans are on the pricing page. If your signals come from TradingView instead of an agent, a webhook bot can run on paper in the same way.
When to consider going live
There is no number of days that makes a strategy safe. These are the conditions worth waiting for:
- The backtest is graded A to C and you have read its findings.
- The overfit verdict is ok, or a walk-forward of the same code came back robust.
- The paper trades match a backtest of the same period, entry by entry.
- You have seen the bot go through at least one losing streak without changing it.
- You know the worst drawdown in the backtest and have decided, in money, that you can sit through it.
Going live is your decision and cannot be made from the chat. On OptAlgo you switch a strategy live in the app, on a Plus or Pro plan. Your agent can suggest it, and only you can confirm it. Live bots run inside limits the server enforces: allocation, leverage and a daily loss limit.
Start smaller than the backtest assumed. The first live weeks are one more test, this time of fills and of your own nerves, the two things paper could not measure. Keep comparing live trades with the backtest in the same way, and treat a growing gap as a reason to go back to paper.
Frequently asked questions
What is a paper trading bot?
A paper trading bot is a trading bot that runs on live market prices with simulated money. It sends no orders to an exchange. Every decision and fill is recorded, so you can check how the strategy behaves before real money is involved.
Can I paper trade crypto with ChatGPT or Claude?
Yes, through a connector. ChatGPT or Claude writes and backtests the strategy code through the OptAlgo connector, and OptAlgo runs it as a paper bot at every bar close. The connector is paper-only and can never place a live order.
How long should I paper trade before going live?
Long enough to see dozens of trades and at least one drawdown. A strategy on short bars can get there in a couple of weeks. A 4-hour trend strategy that trades about once a month needs months, so combine paper with out-of-sample tests such as a walk-forward.
Are paper trading results realistic?
They are realistic about timing and prices for small orders, and optimistic about everything your real order would change. Paper orders do not move the market, are never partly filled and carry no emotional pressure. Expect live results to be somewhat worse.
Is paper trading free on OptAlgo?
The Free plan includes one hosted strategy on paper and 200 paper trades a month, with no card. Paper trading does not use backtest units.