---
title: How many candles for pattern matching? Chart length matters
url: https://comparand.xyz/blog/chart-window-length-matters
description: We changed chart length from 24 to 48 to 96 hourly candles for Bitcoin and Ethereum. Most past matches changed, and no length won everywhere.
updated: '2026-10-21'
category: Patterns
---

# How many candles for pattern matching? Chart length matters

We changed chart length from 24 to 48 to 96 hourly candles for Bitcoin and Ethereum. Most past matches changed, and no length won everywhere.

There is no single best number of candles for pattern matching. In our tests on Bitcoin and Ethereum, changing the chart from 48 to 96 hourly candles kept only **15.20% to 19.57%** of the ten closest past matches, depending on the market. That is about 2 of 10. And no length, whether 24, 48 or 96 candles, gave the best later results in all four markets we tested.

So the chart length is not a small detail. It decides which past cases you see.

For all five search choices step by step, see [How to find past crypto charts that look like today's](https://comparand.xyz/blog/find-past-charts-like-today).

## What does chart window length mean?

A candle, also called a bar, shows the open, high, low and close price over a set time. Two settings shape a chart search:

- **Timeframe:** how much time each candle covers (for example, 1 hour).
- **Window:** how many candles the search uses (for example, 48).

So 48 one-hour candles cover two days. 96 cover four days.

A chart analog is a past chart with a similar shape. Our [plain guide to chart analogs](https://comparand.xyz/blog/what-is-a-crypto-chart-analog) explains how the matching works.

Why would length matter? A short window might show only a drop and a bounce. A longer window ending at the same time also shows the climb before the drop. That is a different shape, so the search looks for different past charts.

*Diagram: Show the same hourly chart ending at a fixed time with windows of 24, 48, and 96 candles, highlighting the older prices added by each longer window.*
## How many past matches stayed the same?

We ran the same 300 search times per market with windows of 24, 48 and 96 hourly candles. The chart's end time stayed fixed; a longer window only added older prices. Each search returned its ten closest past cases. A case counted as "kept" if both searches picked a past chart ending at exactly the same hour.

We tested four markets. Binance BTCUSDT and ETHUSDT are perpetual futures (perps), contracts that follow the coin's price and never expire. Coinbase BTC-USD and ETH-USD are spot markets, where the coins themselves trade. Every search compared a coin only with its own past on the same exchange.

| Market (hourly, searches from Jan 2020 to Sep 29, 2026) | 24 vs 48 candles | 48 vs 96 candles | 24 vs 96 candles |
| --- | ---: | ---: | ---: |
| Binance BTCUSDT | 15.57% kept | 19.57% kept | 5.67% kept |
| Binance ETHUSDT | 14.47% kept | 18.97% kept | 5.53% kept |
| Coinbase BTC-USD | 12.77% kept | 18.80% kept | 4.43% kept |
| Coinbase ETH-USD | 11.80% kept | 15.20% kept | 4.00% kept |

Each row is 300 searches. Binance ETHUSDT searches start on January 18, 2020; the others start on January 1, 2020.

In plain numbers: doubling the window usually kept about 1 or 2 of the 10 matches. The typical (median) search kept 1 case between neighbouring lengths and 0 between 24 and 96 candles.

Often nothing was kept at all. Between 24 and 96 candles, the two lists shared no case in 69.00% of Binance Bitcoin searches and 76.33% of Coinbase Ethereum searches.

## Could these be nearby hours of the same price move?

Matching the exact same ending hour is strict. A longer search might pick a past chart ending a few hours earlier or later, during the same move.

So we also let each case pair with one case from the other list ending within 12 hours. This barely helped. For Binance Bitcoin, 48 vs 96 candles went from 1.96 to 2.04 kept cases out of 10. The lists were really different, not just shifted by a few hours.

We ran one more check. A shorter window can reach a bit closer to the present, because its past charts are shorter. So we gave every window the same latest allowed past ending. The results barely moved (for Binance Bitcoin, 48 vs 96 stayed at 1.96 of 10).

One limit stays: longer windows need more space between matches so their charts do not overlap. Our results include that effect along with the change in shape.

## Did a longer chart give better results?

No. We checked what happened in the 24 hours after each search. For each search, we took the middle (median) later move of its ten matches, and pretended to trade in that direction. This was a research calculation to compare lengths, not a suggested trade. We charged 0.05% to enter and 0.05% to exit.

| Market | 24 candles | 48 candles | 96 candles |
| --- | ---: | ---: | ---: |
| Binance BTCUSDT | −0.04% | −0.30% | −0.19% |
| Binance ETHUSDT | +0.07% | +0.16% | +0.09% |
| Coinbase BTC-USD | −0.06% | −0.25% | +0.07% |
| Coinbase ETH-USD | −0.69% | −0.01% | +0.24% |

These are average returns per search after fees, over 300 searches per market.

The winner changed from market to market. 96 candles looked best on Coinbase. 48 looked best on Binance Ethereum. On Binance Bitcoin, every length lost on average.

We also counted how often the matches pointed the right way over the next 24 hours. Across all lengths and markets, this stayed between 42.00% and 55.00%. For comparison, the price simply went up after 47.33% to 54.33% of the same search times, depending on the market. So no length gave a clear edge in this test.

One thing did not change with length: the ten matches often disagreed. In 63.33% to 70.67% of searches, only 4, 5 or 6 of the ten went up over the next 24 hours, for every length and market. We look at those split answers in [how often similar charts disagree](https://comparand.xyz/blog/when-past-crypto-cases-disagree).

If you tried all three lengths and kept the best one for each market, you would be fitting your settings to luck. Our [overfitting test](https://comparand.xyz/blog/overfitting-1000-random-trading-rules) shows how that goes wrong.

## How should you pick a chart length?

Here is a practical way to choose, based on what we saw:

1. **Decide what move you care about first.** If you care about the last day's drop, 24 hourly candles may be enough. If you care about the week before it, you need more.
2. **Pick your length before looking at outcomes.** Write it down.
3. **Treat a new length as a new search.** Since 80% or more of matches changed, do not expect the same story.
4. **Check two or three lengths you chose in advance.** If a pattern in the later moves only shows up at one length, be careful with it.
5. **Compare the same waiting period.** Look at the same later horizon (for example, 24 hours) for every length you compare.

If you have no reason to pick a length, Comparand's chart search starts at 48 candles. That is a starting point, not a tested best: 48 did not win in our test either.

## How we measured this

We used public Binance and Coinbase hourly prices, with searches spread from January 2020 to September 29, 2026, 300 per market. For each search, we turned the last 24, 48 or 96 hourly price changes into a shape on a common scale and found the ten closest past charts from the same coin on the same exchange, with no overlapping charts. Every past match had to have its full 48 hours of later prices already known at the search time, so no future data was used. The test trades charged 0.05% per side and left out funding, borrowing, spread, slippage and delays; a short on a spot market would need borrowed coins. Missing hours were not filled in, and the same past cases can appear in many searches, so results are not independent. Four markets cannot prove the best length for other coins or periods.

## Try this in Comparand

You can run this comparison yourself with the chart search, shape_analogs. Ask your AI agent:

**"Find ten past Binance Bitcoin charts like the hourly chart ending at [your time]. Run it with windows of 24, 48 and 96 candles, and show which past endings appear in more than one list."**

What comes back for each window:

- Up to 10 past cases (you can ask for 1 to 20) from Bitcoin's own Binance past, on the same timeframe, never overlapping your chart.
- Each case's own price path and its later moves at a quarter, half and the full window length, in basis points (one basis point is one hundredth of one percent).
- Individual cases, with no average or hit rate. Your agent can compare the lists.

Settings you control:

- **Window:** 20 to 200 completed candles (default 48).
- **Timeframe:** 5 minutes, 15 minutes, 1 hour (default), 4 hours or 1 day.
- **End time:** any past time, or now.

Note that the later moves are reported at a share of the window. With 24 candles, the first one is 6 hours later; with 96, it is 24 hours later. Compare the same horizon across lengths.

Each search costs $0.01 from credits. The [docs](https://comparand.xyz/docs) explain the settings. You can [try 3 coins free](https://comparand.xyz/try). These are past cases, not a forecast.

## The short version

- Going from 48 to 96 hourly candles kept only 15.20% to 19.57% of the ten closest past matches in our Bitcoin and Ethereum tests.
- From 24 to 96 candles, the lists often shared no match at all.
- No length gave the best results in every market, and none gave a clear edge.
- Choose a length that fits the move you care about, before you look at outcomes.
- Treat each new length as a new search, and compare outcomes over the same horizon.

## Questions people ask

**How many candles should I use for pattern matching?**

There was no single best number in our tests. Across Bitcoin and Ethereum on Binance and Coinbase, no window length of 24, 48 or 96 hourly candles gave the best result in every market. Pick the length that covers the move you care about, and pick it before you look at outcomes.

**Does a longer chart find the same past matches?**

Mostly not. In 300 hourly Binance Bitcoin searches from 2020 to September 2026, going from 48 to 96 candles kept 19.57% of the ten past matches on average. Going from 24 to 96 kept 5.67%.

**Is window length the same as timeframe?**

No. The timeframe is how much time each candle covers, such as 1 hour. The window length is how many candles the search uses, such as 48.

**Did longer charts give more accurate matches?**

No. In our tests, how often the matches pointed the right way over the next 24 hours stayed between 42% and 55% for every length and market, close to how often the price simply went up.

**Does Comparand turn the matches into a prediction?**

No. The chart search shows each past case and what happened after it, with no average or hit rate. The averages in this article come from our own separate test.