---
title: How many trades does a backtest need to be significant?
url: https://comparand.xyz/blog/how-many-trades-backtest-sample-size
description: Luck moves a 50% hit rate about 10 points at 100 trades and 3 at 1,000. Why your trades may not be separate, and what 3,600 real tests could not show.
updated: '2026-10-21'
category: How-to
---

# How many trades does a backtest need to be significant?

Luck moves a 50% hit rate about 10 points at 100 trades and 3 at 1,000. Why your trades may not be separate, and what 3,600 real tests could not show.

There is no single number that is always enough. For a hit rate near 50%, luck alone moves the result about 10 points either way at 100 trades and about 3 points at 1,000, so an edge has to be bigger than that to stand out. In our test of chart look-alikes, 3,600 tries picked the next-day direction 50.22% of the time, against 50.92% for always guessing "up". That 0.70-point gap sits well inside the normal wobble of under 2 points, so the data could not say which was better. These are past cases, not a forecast.

For the full chart look-alike test used here, see [Do chart patterns predict crypto prices?](https://comparand.xyz/blog/do-chart-lookalikes-predict-crypto).

## Why does the number of trades matter?

A hit rate is the share of trades that went the right way. If a rule has no real edge, its hit rate still will not land exactly on 50%. It wobbles around 50%, like a coin flipped many times. Few flips wobble a lot. Many flips wobble less.

A good rule of thumb for a hit rate near 50%: the normal wobble (the 95% margin, meaning about 19 times in 20 a no-edge rule lands inside it) is about 1 divided by the square root of the number of trades, as a fraction.

| Number of trades | Normal wobble around 50% (percentage points) |
| --- | --- |
| 30 | about 18 points either way |
| 100 | about 10 points |
| 300 | about 6 points |
| 1,000 | about 3 points |
| 3,600 | under 2 points |
| ten thousand | about 1 point |

To cut the wobble in half, you need four times as many trades. You can also read the table the other way when you plan a test: a 10-point edge needs about 100 separate trades before it stands out from luck, a 6-point edge about 300, and a 3-point edge about 1,000.

*Diagram: The normal wobble band around a 50% hit rate shrinks from about 18 points at 30 trades to under 2 points at 3,600, with the 12 real market hit rates at 300 tries each and the pooled 50.22% at 3,600 marked on it.*
## What did 3,600 tests show?

We built a test with 12 markets: 10 Binance perps (futures contracts that never expire) and 2 Coinbase spot markets. For each, we picked 300 evenly spaced dates from January 2020 to September 2026. On each date we found the 10 closest past charts for the same coin on the same exchange, and called "up" or "down" from their middle 24-hour result.

Across all 3,600 tries, the chart-matching guess was right 50.22% of the time, and always guessing "up" 50.92%. After an assumed 0.05% fee on each side, the average 24-hour trade lost 0.07%. We cannot say either approach is better. We can say chart matching showed no edge here (full story: [do chart look-alikes predict crypto?](https://comparand.xyz/blog/do-chart-lookalikes-predict-crypto)).

## What does a small sample look like?

Now split the same test by market. Each market has only 300 tries, so the wobble is about 6 points. Here is each market's 24-hour hit rate, with always guessing "up" beside it.

| Market | Chart-match hit rate | Always guessing "up" |
| --- | --- | --- |
| Binance BNBUSDT | 43.67% | 51.00% |
| Coinbase BTC-USD | 47.33% | 49.00% |
| Binance ADAUSDT | 48.67% | 47.33% |
| Binance LINKUSDT | 49.00% | 55.67% |
| Binance XRPUSDT | 50.33% | 49.33% |
| Coinbase ETH-USD | 50.33% | 47.33% |
| Binance BTCUSDT | 50.67% | 49.33% |
| Binance DOGEUSDT | 50.67% | 52.67% |
| Binance ETHUSDT | 51.33% | 54.33% |
| Binance TRXUSDT | 52.33% | 54.67% |
| Binance SOLUSDT | 53.67% | 49.67% |
| Binance AVAXUSDT | 54.67% | 50.67% |

The hit rates run from 43.67% to 54.67%. Someone who only looked at Avalanche might think the method works. Someone who only looked at BNB might think it fails. Yet one method, which showed no edge overall, produced both. That is what 300 coin flips look like.

In the chart above, BNB's 43.67% is the only market just outside the normal wobble band for 300 tries. That is not surprising: the band holds about 19 in 20 no-edge results, so with 12 markets, one stray is normal.

The "up" column also moves, from 47.33% to 55.67%. A baseline is a measurement too, with its own wobble.

## Are your trades really separate?

Counting trades is not the same as having separate tests. Three things shrink your real sample.

- **Overlapping windows.** In a separate study, we measured 24-hour moves from every hourly start: 62,335 cases for Binance BTCUSDT. A start at 1:00 and one at 2:00 share 23 of their 24 hours, so that is closer to one fresh day per 24 starts than to 62,335 separate tests.
- **Reused history.** In the 3,600 tests, matches came from the same past, and Bitcoin alone reused some of the same past endings. So the tests are linked.
- **Markets that move together.** The 12 markets rise and fall in similar periods. Twelve markets do not give you twelve times the evidence.

Within each market, our test dates were more than 48 hours apart, so no two tests shared the same 24 hours. But the other two effects remain, so we make no claim about statistical significance from the count alone.

## What changes when you try many rules?

The wobble table assumes one rule, tested once. If you try many, the best one will look good by luck.

We tested this. We made 1,000 random rules on Binance BTCUSDT hourly prices. Each looks back a random number of hours, uses a random size threshold, and either follows or fades the move. We picked the best on training dates and froze it for later dates. Sharpe is average return divided by how much returns bounce around, scaled to a year (see [Sharpe ratio explained](https://comparand.xyz/blog/sharpe-ratio-explained-crypto)).

| Result for the rule picked as best | Training: Sep 2019 to 2022 | Later dates: 2023 to Sep 2026 |
| --- | --- | --- |
| Hours of data | 29,046 | 32,856 |
| Sharpe | 0.75 | −0.19 |
| Return after fees | +155.22% | −51.72% |
| Worst drop | −60.22% | −76.83% |

The 29,046 training hours sound like plenty. It did not help. Only 24.5% of the 1,000 rules had a positive training Sharpe, and the best one was the luckiest of the group. On later dates, only 1.00% had a positive Sharpe, and 83.6% did worse than in training.

So the more rules you try, the more proof the winner needs. Judge the best of 1,000 against what the best of 1,000 random rules usually scores, then test it on dates it never saw ([walk-forward testing explained](https://comparand.xyz/blog/walk-forward-testing-explained)). For the full test, see [overfitting with 1,000 random trading rules](https://comparand.xyz/blog/overfitting-1000-random-trading-rules).

## How do the forum rules of thumb hold up?

- **"30 trades is enough."** No. The wobble is about 18 points either way. Only a huge edge would show.
- **"A few hundred trades."** Better. At 300 the wobble is about 6 points. You could see a 12-point edge, but not a 2-point one, and our market-by-market table shows how noisy it still is.
- **"1,000 trades."** The wobble is about 3 points. That helps if you tried one rule and the trades are separate. It does not help if you picked the best of many.
- **"Years of data."** Years help only through the separate tests inside them. See the overlapping windows point above.

A hit rate is also not a return. A single day often swings several percent either way. For Binance BTCUSDT, the middle 80% of 24-hour moves ran from −3.07% to +3.49%. To see a tiny average gain through swings like that, you need thousands of trades.

## How we measured this

For the chart-match test, each market's 300 decisions were matched against its own earlier windows of 48 hourly closes. We kept the 10 closest non-overlapping matches and scored direction 24 hours later against what really happened. The 1,000-rule test used Binance BTCUSDT hourly bars from September 2019, a 0.05% fee on each side and no funding costs, with training through 2022. The wobble numbers are standard coin-flip math and assume separate trades, which these were not fully.

## Try this in Comparand

You can run your own repeated-test study on Hyperliquid data, separate from the Binance and Coinbase numbers above.

- Ask your agent: "Call hl_coverage, then run hl_analogs_batch for these decision times for ETH with a 1h horizon." **hl_analogs_batch** takes up to 100 queries in one call, each with its own record of its data sources, as JSON or CSV. It costs $0.01 for each search that runs; refused searches are not charged.
- It returns the nearest past moments and what followed each one, not an average or a hit rate. Your agent does the counting, so you control how many tests you run and how you score them.
- Ask: "Use hl_bulk_files, then hl_bulk_link for the days I choose." **hl_bulk_files** is a free catalog of whole archive day files. **hl_bulk_link** gives signed download links for 1 to 100 days of one table, each good for 24 hours and 5 downloads. Plans include a monthly amount of link data; pay as you go costs $0.15 per GB.

The Hyperliquid archive is shorter than the Binance history above, so check how many days it holds before you count tests. You can also [try 3 coins free](https://comparand.xyz/try), and see [pricing](https://comparand.xyz/pricing) for plan limits.

## The short version

- For a hit rate near 50%, 100 trades leave about 10 points of normal wobble. 1,000 trades leave about 3.
- 3,600 tests could not separate 50.22% from 50.92%, and chart matching showed no edge.
- With 300 tries per market, hit rates ran from 43.67% to 54.67% with no real edge.
- Overlapping windows, reused history and linked markets make the true sample smaller than the count.
- The more rules you try, the more proof the winner needs; the best of 1,000 random rules lost 51.72% on later dates.

## Questions people ask

**How many trades does a backtest need to be significant?**

There is no single number. For a hit rate near 50%, 100 trades leave a normal wobble of about 10 points either way, and 1,000 trades about 3 points. The trades also need to be separate, and trying many rules needs more proof than trying one.

**Is 30 trades enough for a backtest?**

No, not for a hit rate near 50%. With 30 trades the normal wobble is about 18 points either way, so a 60% result can be luck. Only a huge difference would stand out.

**Why could 3,600 tests not tell 50.22% from 50.92%?**

The gap is 0.70 points. With 3,600 coin-flip-like tests, the normal wobble is under 2 points either way, so a gap that small is inside it. The data cannot say which is better.

**Does more data fix overfitting?**

No. In our test, the best of 1,000 random Bitcoin rules had 29,046 hours of training data and a +155.22% result after fees. On later dates it lost 51.72%. Many hours of data did not protect against picking a lucky rule out of many.

**How many trades do you need to spot a small edge?**

For a hit rate near 50%, a 10-point edge needs about 100 separate trades to stand out from luck, a 6-point edge about 300, and a 3-point edge about 1,000. Smaller edges need thousands.