Build and test a strategy Lesson 13 of 14
Strategies · Building and testing a strategy
Paper trading a strategy: what a practice record proves
What paper trading can and cannot show about a strategy, how many trades a record needs, and why the first 30 trades of a real rule pointed the wrong way.
On this page
Paper trading means following a strategy with pretend money at real prices. It sits between a backtest and a real account. A backtest shows what a rule would have done on data that already exists. Paper trading shows what it does on data that did not exist when the rule was written, with the person in the loop: reading the signal, placing the order, keeping the record.
It is also widely misused. People paper trade for a month, see a profit and conclude the strategy works, or see a loss and drop it. A month of trades can show neither. This lesson sets out what a paper record can prove, what it cannot, and how long it has to be, using 759 real trades of one simple rule to show how misleading a short record is. How to place paper orders on Finabase is in Practise with Finabase paper trading on the web.
Where paper trading fits
A strategy that deserves real money usually passes through three stages.
- Backtest. Write the rules, test them on history with costs, split the period, try the neighbouring settings (Backtesting basics: biases, costs and an honest SPY test). This is cheap and fast, and it is where most ideas should die.
- Paper trade. Follow the rules forward in real time with virtual money. Nothing about the rules may change during this stage; a changed rule starts a new record.
- Trade small. Use real money at a size where the loss does not matter, to find out what the first two stages could not show.
Each stage removes a different kind of error. The backtest removes ideas that never worked on any data. Paper trading removes rules that only worked on the data they were built from, and rules that cannot be executed as written. Small real trading reveals the costs and the behavior.
What a paper record can show
Whether the rules are complete. A backtest program never asks "what do I do now?". A person following rules by hand asks it within days. What if the signal comes on a day the trade cannot be placed? What if two signals arrive at once and there is cash for one? What if the stock gaps past the entry price? Each of these is a missing rule, and paper trading finds them for free.
Whether the rules can be carried out. Many backtests trade at the closing price of the bar that gave the signal. To do that in practice, the signal has to be calculated a few minutes before the close from a price that is not yet final. Paper trading shows whether that is workable, and how often the nearly-final signal differs from the final one.
Whether the result survives new data. This is the important one. Every trade taken after the rules were fixed is out of sample. The rule could not have been fitted to it. If a strategy's paper results look like its backtest, that is real evidence. If they do not, the backtest was probably fitted to its history.
Whether the trader follows the rules at all. Even with no money at stake, people skip signals they dislike and take trades the rules did not give. A paper record that marks each trade "rule" or "not rule" measures this directly (Trading plan and trading journal: what to write down).
What it cannot show
Real fills. A paper order is filled at a price taken from the live market, as if the order had not been there. A real order has to find someone to trade with. A market order pays the spread and, if it is large relative to the shares on offer, moves the price. A limit order waits in a queue and may never be filled, and it is most likely to be filled exactly when the price is about to go through it. Paper results are therefore an upper bound, and the gap is widest for strategies that trade often, trade small or thinly traded stocks, or trade at the open and close (Liquidity and slippage: volume, depth, spread and market impact).
Behavior under real losses. Sitting through a paper drawdown costs nothing. The decisions that damage real accounts, selling at the low, doubling a losing position, abandoning a rule after three losses, are made under a stress that virtual money does not produce (Trading psychology: loss aversion, overconfidence and other biases).
Rare events. A few months of paper trading samples one kind of market. It will not contain a crash, a trading halt or a year-long decline unless one happens to occur.
How many trades are enough?
This is the question a paper trader most needs answered, and it has a numerical answer.
The uncertainty of a win rate
If a strategy wins a share p of its trades, the win rate measured over n trades will vary from one sample to the next. Its standard error is:
A range of two standard errors either side covers the truth about 95 times in 100. So after 30 trades, a measured win rate of 55% is consistent with a true rate anywhere from about 37% to 73%. After 100 trades the range is 45% to 65%, which still includes a coin flip. Only at around 400 trades does the range, 50% to 60%, stop including 50%.
The error shrinks with the square root of the number of trades. To halve the uncertainty takes four times the trades.
The same thing on real trades
To see this happen, take one simple short-term rule and collect every trade it made. The rule, used here only as a source of trades: buy at the close when the two-day RSI is below 10; sell at the first close above the five-day simple moving average.
- Data. Daily bars from November 18, 2024 to November 17, 2026, 501 trading days, for each of the 30 stocks in the Dow Jones Industrial Average.
- Trading and costs. Trades at the close of the signal day; 0.10% of the amount on each buy and each sell.
- Left out. Dividends and taxes. The stocks are the index's members at the end of the period.
The rule made 759 completed trades across the 30 stocks. Put them in the order they closed, as a paper record would have grown, and track the win rate as it goes.
| Trades recorded | Ended | Win rate so far | Two standard errors | Average gain per trade |
|---|---|---|---|---|
| First 20 | December 2024 | 45.0% | ±22.2 points | −1.03% |
| First 30 | December 2024 | 46.7% | ±18.2 points | −1.56% |
| First 50 | January 2025 | 60.0% | ±13.9 points | −0.77% |
| First 100 | March 2025 | 60.0% | ±9.8 points | −0.42% |
| First 300 | September 2025 | 64.7% | ±5.5 points | 0.17% |
| All 759 | November 2026 | 65.6% | ±3.4 points | 0.36% |
After 30 trades, a month of paper trading, the rule had won fewer than half its trades and lost 1.56% on the average trade. Anyone judging it then would have dropped it. After 100 trades, four months in, the average trade was still a loss. Over all 759 trades it won 65.6% and gained 0.36% per trade after costs.
The first 30 were not a fluke of ordering. Cut the 759 trades into 25 consecutive blocks of 30: the average gain per trade in a block ranged from −1.74% to +2.56%, seven of the 25 blocks lost money, and the win rate of a block ranged from 43% to 87%. A 30-trade record of this rule could have said almost anything.
Enough trades for the average, too
A win rate is not a profit. What matters is the average gain per trade, and its uncertainty depends on how widely the results are spread:
For these trades the average was 0.36% and the spread (the standard deviation of the individual results) was 3.19%. That gives (2 × 3.19 / 0.36)², about 320 trades. A strategy with a larger average relative to its spread needs fewer; most need more than people expect.
Two cautions make even this figure optimistic. Trades on 30 stocks over the same days are not independent, because the stocks fall together, so 759 trades carry less information than 759 separate experiments. And the two years were one kind of market. Splitting the record in half by date gives an average of 0.22% per trade up to December 2025 and 0.49% after, which is a reminder that even hundreds of trades describe only the period they came from.
One symbol looks better than it is
On SPY alone over the same two years the rule made 29 trades, won 79% of them and returned 26.5% against 24.5% for holding the fund, while invested only 16% of the time. That looks excellent. It is 29 trades. The two-standard-error range on that win rate is ±15 points, and across the 30 stocks the same rule lost money on 8, from a gain of 49.8% on the best to a loss of 38.6% on the worst. A paper trader who happened to pick the best symbol would have learned the wrong lesson as surely as one who picked the worst.
What the fill assumption does
Because paper fills are generous, the cost charged on paper should be chosen on purpose. The same 759 trades at different assumed costs per buy and per sell:
| Cost per trade (each way) | Average gain per trade | Win rate |
|---|---|---|
| None | 0.56% | 68.6% |
| 0.05% | 0.46% | 67.1% |
| 0.10% | 0.36% | 65.6% |
| 0.20% | 0.16% | 63.1% |
| 0.30% | −0.04% | 59.8% |
A rule that holds for a few days earns a fraction of a percent per trade, so a few hundredths of a percent of slippage each way is a large share of it. At 0.30% each way the profit is gone and the win rate is still nearly 60%, which is one more reason not to judge by win rate. For large, heavily traded stocks during the trading day 0.30% is pessimistic; for small stocks, or orders at the open, it is not.
The practical step is to record, for each paper trade, the price the rules assumed and a deliberately worse price, and to keep both totals.
A paper-trading plan
- Write the rules in full before the first trade, including size, and date the page.
- Work out from the backtest how often the rule trades and how many trades the formulas above call for. Set the length of the paper period from that number, not from the calendar.
- Fix the cost assumption and apply it to every trade.
- Take every signal. Mark any trade taken or skipped against the rules.
- Do not look at the running total every day. Review at the planned number of trades.
- Compare with the backtest on the things that should match: trades per month, win rate, average win and loss, time in the market. Differences in those are more informative than the profit.
- Change nothing mid-record. A new rule is a new record starting from zero.
Where paper trading misleads
- Stopping when ahead. Ending the test at a high point is choosing the sample after seeing it. The stopping point has to be fixed first.
- Restarting after losses. Resetting the account when it is down and beginning again keeps only the good runs.
- Sizes that will not be used. A $10,000 paper position in a stock that trades 20,000 shares a day teaches nothing about a real order there.
- Slow strategies. A rule that trades once a year cannot be paper traded to a conclusion in any reasonable time. For those, the long backtest and a small real position are the only evidence available, and paper trading serves only to check that the rule can be followed.
How Finabase does it
Finabase's paper trading is an account with a one-time $10,000 of virtual cash (Practise with Finabase paper trading on the web). It accepts buy and sell orders, market or limit, good for the day or until cancelled. Orders fill on the next matching trade in the live market, so prices are real and the fill assumption is the generous one described above: a paper market order does not move the price or wait in a queue. The account shows value, cash, realized and unrealized gain, a history chart and the full transaction history, which is the raw record this lesson asks for. It does not mark trades as following a rule or not, and it does not charge commissions or slippage, so keep those two columns yourself.
The two research tools answer a different question from paper trading (Backtest indicator signals and rule-based strategies on Finabase). Backtest runs indicator signals over past data, buying or shorting a fixed $10,000 per signal with no trading cost, and is the place to count how often a rule trades before deciding how long to paper trade it. Strategy replays if/then rules over the last two years. Both look backward; only the paper account produces trades the rules could not have been fitted to. Resetting the paper account removes all its orders, so decide the length of a test before starting, not after a bad week.
Frequently asked questions
How long should a strategy be paper traded?
Until it has made enough trades, which is a count and not a length of time. For a rule with a small average gain per trade that is typically several hundred trades. A rule trading twice a week needs years; one trading ten times a day needs weeks. A record of 30 trades cannot distinguish a good rule from a bad one.
Is paper trading accurate?
The prices are real and the fills are optimistic. A paper order does not pay for its own effect on the market or miss fills the way a real limit order can. Treat paper results as the best case and subtract a deliberate allowance for costs.
Why did the strategy work on paper and lose with real money?
The usual causes are a paper record too short to mean anything, fills that were better on paper than in the market, and decisions made differently once losses were real. A short lucky record is the most common of the three.
Does paper trading replace backtesting?
No. A backtest covers decades in seconds and paper trading covers months in months. Use the backtest to reject ideas and paper trading to confirm that a surviving idea behaves on new data as it did on old.
Summary
Paper trading tests a strategy on data it could not have been fitted to, and shows whether its rules are complete and can be followed on time. It does not reproduce real fills or real nerves. Its results mean something only when there are enough trades: the uncertainty of a win rate after 30 trades is about ±18 points, and in the example a rule that won 66% of 759 trades won 47% of its first 30 and lost money on its first 100. About 320 trades were needed before that rule's average gain could be told from zero, and raising the assumed cost from 0.10% to 0.30% each way removed the profit entirely. Decide the number of trades and the cost assumption first, take every signal, and compare the record with the backtest when the count is reached.
Check your understanding
5 questions on this lesson. Each answer comes with the reason, and nothing is scored.
After 30 paper trades a strategy has won 55% of the time. What can be said about its true win rate?
- It is close to 55%
- It is somewhere between roughly 37% and 73%
- It is above 50%, so the strategy has an edge
Show the answer
B. It is somewhere between roughly 37% and 73% The standard error of a 55% win rate over 30 trades is about 9 points, so two standard errors either side runs from about 37% to 73%. That range includes a coin flip and a strongly losing rule. About 400 trades are needed before it excludes 50%.
A rule lost money on its first 100 paper trades and was profitable over 759. What is the lesson?
- Rules improve with practice
- The first 100 trades must have had data errors
- A record of that length is too short to judge the rule either way
Show the answer
C. A record of that length is too short to judge the rule either way In blocks of 30 trades the same rule’s average gain ranged from −1.74% to +2.56% and seven of 25 blocks lost. Short records of a noisy rule can point in either direction, so the number of trades has to be fixed before the test begins.
Which of these can paper trading NOT show?
- Whether the rules cover every situation that comes up
- Whether a large order would have been filled at the quoted price
- Whether the rule behaves on new data as it did in the backtest
Show the answer
B. Whether a large order would have been filled at the quoted price A paper order is filled as if it had no effect on the market and no queue to wait in. Real orders pay the spread, can move the price and can go unfilled. Completeness of the rules and behavior on new data are exactly what a paper record is for.
The same trades averaged +0.36% each at a cost of 0.10% each way and −0.04% at 0.30%, while the win rate stayed near 60%. What follows?
- Costs do not matter when the win rate is high
- A short-term rule’s profit is small per trade, so the fill assumption can decide the result
- The rule should be traded more often to spread the cost
Show the answer
B. A short-term rule’s profit is small per trade, so the fill assumption can decide the result A rule that holds for days earns a fraction of a percent per trade. A few hundredths of a percent of slippage each way is a large share of that, which is why paper results, with their generous fills, are a best case.
A paper account is down after three weeks, so the trader resets it and starts again. What is wrong with that?
- It keeps only the good runs, so the eventual record is selected after the fact
- Resetting changes the prices the orders fill at
- Nothing; early losses are not representative
Show the answer
A. It keeps only the good runs, so the eventual record is selected after the fact Stopping when ahead and restarting when behind both choose the sample after seeing it. The length of the test, in trades, has to be decided before the first one and kept.
Exercise
One task to do in Finabase. Nothing is graded or saved.
Open Paper trading and sign in (a free account; press Fund my virtual account on the first visit). During market hours, write down the SPY price you see, then place a paper Market order to buy 1 share. Find the fill in Transaction history. Compare the fill price with the price you wrote down, then write a second, deliberately worse price 0.10% higher. Which of the two would a real order be closer to?
Open paper tradingWhat you should find
The paper fill is taken from the live market as if your order were not there, so it lands very close to the price you saw. A real market order pays the spread and can move the price, so the worse column is the more realistic one, and paper results are an upper bound. Keep both columns for every trade: the account does not charge costs or slippage for you.
Sources and further reading
- SEC Office of Investor Education and Advocacy, “Investor Bulletin: Performance Claims” — on hypothetical and back-tested results
- Brad M. Barber and Terrance Odean, “Trading Is Hazardous to Your Wealth,” Journal of Finance 55(2), 2000 — how individual investors actually trade
- David H. Bailey and Marcos López de Prado, “The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality,” Journal of Portfolio Management 40(5), 2014
- Finabase chart data — daily bars of SPY and the 30 Dow Jones Industrial Average stocks, November 18, 2024 to November 17, 2026; rule and results computed by Finabase
