Why Short-Term Crypto Price Prediction Is Hard: The Coin-Flip Baseline and Hit-Rate Illusions

The coin-flip baseline

Calling whether bitcoin will be higher or lower in 5 minutes or an hour looks easy. But picking 'up' or 'down' without any thought gives a hit rate around 50%. So any prediction method must first be compared with this coin-flip baseline. Measured on real data, the baseline may not be exactly 50%. Using this site's chart-course statistics script on 30,076 Binance spot daily candles for 10 coins (BTC, ETH, BNB, XRP, ADA, DOGE, LTC, LINK, TRX, SOL; 2017-08-17 to 2026-10-06, measured on 2026-10-07), the next day's close was higher 50.6% of the time. For bitcoin's daily candles alone, the close five days later was higher 53.2% of the time (3,333 samples). Because coins rose a lot over this period, always answering 'up' would have been right that often. It is fairer to set the baseline not at 50% but at 'the hit rate from always answering up'.

The random walk explanation

If short-term price changes happen almost independently of the change just before, it is very hard to predict the next move from the shape of the past chart alone. Such movement is called a random walk, and it is widely used as a starting point for describing prices over short spans. That does not mean markets are perfectly random. But if there were a rule anyone could see on a public chart, many people would act on it first and it would tend to fade quickly. When this site's chart course measured patterns directly on past Binance candles, most candlestick patterns were followed by rises at a rate within a few percentage points of 'any candle'. The figure below is a price made only by flipping a coin. It contains no information at all, yet shapes that look like trends, bottoms and rebounds appear.

Hit rates cannot be trusted on small samples

Getting 12 of 20 predictions right is 60%. That seems good, but pure coin flipping gets 11 or more of 20 right about 41% of the time. With few samples, high hit rates often arise by chance alone. With 30 calls, roughly 32 to 68% is common by chance; with 100 calls, 40 to 60% is common. This range narrows in proportion to the square root of the number of calls, so telling a small edge such as 55% apart from chance takes roughly 400 independent calls. And consecutive predictions are often not independent of each other, so in practice you need an even longer record.

• 11 or more of 20: about 41% by chance
• 30 calls: about 32 to 68% is common by chance
• 100 calls: about 40 to 60% is common by chance
• To tell 55% from 50%: roughly 400 independent calls

Survivorship illusion: you only see those who called it right

Suppose 1,000 people each flip a coin to make 10 predictions. On average about 1 person (1,000 ÷ 1,024) gets all 10 right, and about 55 get 8 or more right. Even with no skill, dozens of people end up with 'a hit rate of 80% or more', and their records are what people mostly see, because the 950 or so who were wrong quietly disappear. When you see someone's remarkable hit record, ask how many people predicted the same way, where the wrong calls are, and whether the predictions were published in advance or picked out afterwards. This applies to your own record as much as anyone else's, because it is easy to remember the days you were right and forget the days you were wrong.

Expected value matters more than hit rate

Even with a hit rate above the baseline, profit and loss can differ. The result depends on how often you were right, how much you made when right and lost when wrong, and costs. Expected value per trade can be written as hit rate × average gain − (1 − hit rate) × average loss − costs. For example, even with a 60% hit rate, if you make +1% when right and lose −2% when wrong, 0.6 × 1 − 0.4 × 2 = −0.2%, a loss before costs. Conversely, a 40% hit rate can turn out differently if wins are large and losses small. The shorter the time frame, the smaller each move, so fees and spreads take up a larger share of expected value.

Hindsight bias

Looking at a past chart, where the bottom and the top were seems obvious. This makes it easy to feel 'you could have known then', which is called hindsight bias. At that moment the right side of the chart was hidden, and the same shape often went the other way afterwards. If you practice stepping forward one candle at a time with the future hidden, you can feel directly that the clarity you see in hindsight is barely visible in real time. Writing down your reasons before predicting and recording them next to the results lets you later compare, against the baseline, whether those reasons actually helped. Other illusions in reading charts are covered in the guide on the limits of chart analysis.

Using the tools on this site

This site's Live Crypto Prediction Game lets you pick a coin and 5 minutes, 15 minutes or 1 hour, choose up or down, records the start price from Binance's live trade price, and judges the result with Binance's 1-second candle close at expiry. It uses no money or points and always shows the coin-flip range for your number of calls beside your record. The Chart Replay Trainer picks a random stretch of real past Binance charts, hides the coin name, dates and the future, lets you trade virtually as you step forward one candle at a time, and compares your result with simply holding through the same stretch. The Chart Quiz can be used for the same purpose. These tools are better used to honestly measure whether your judgment beats the baseline than as training to build forecasting power. How to keep replay records is in the chart replay practice guide.

Limits and disclaimer

The statistics here were measured on past Binance daily candles with this site's statistics script; measured again, the period grows and the numbers change. Past rates of rising do not mean the same in the future. The probability calculations assume independent coin flips, and the figure is an illustration of the principle. This guide explains why short-term prediction is hard and how to read hit rates. It does not teach a forecasting method and is not a trading signal or investment advice.