Bitcoin Price Prediction

Why a price prediction cannot be known

A price is the result of many people deciding to buy or sell at once. Their choices depend on news that has not been written, on money that has not yet arrived and on moods that change. None of that is in the past data. A model sees only the numbers, so it cannot see most of what moves the price next.

Two more things make it worse. First, a pattern that does appear in the past may be luck. With enough charts and enough rules, some rule will always fit, as the page on overfitting shows. Second, a pattern that is real can fade once people act on it. A rule that worked on old data is a fact about old data.

So the honest answer to "where will bitcoin be next month" is that nobody knows. What can be known is how large the moves have been, how much a position would lose in a bad run, and how a rule behaved when tested fairly. Those are the useful questions, and this site covers them.

How a forecast model is built

Every forecast model has the same three parts.

• Inputs. Past prices, volume, other markets, or text such as headlines. Each input is a choice, and each choice is a setting.
• A rule. A formula, a set of if-then steps, or a machine learning model that turns the inputs into a guess. The guess might be the next price, the next change, or just up or down.
• Fitted values. The model's numbers are chosen so the rule matches past data as closely as possible. That step is called fitting or training.

The more numbers a model can adjust, the more closely it can match any past. A model with enough freedom can match pure noise perfectly. That is why a close match to the past is not evidence. For the machine learning side of this, see AI stock prediction.

How a forecast is tested

The fair test is to split the data by time. Fit on the earlier part. Then score on the later part, which the model never saw. That later part is called out of sample.

The score needs something to be compared with. The simplest baseline is the naive forecast: tomorrow equals today. If a model cannot do better than that, it adds nothing.

Two traps spoil the split. One is look-ahead bias, where information from after the forecast date leaks into the inputs. The other is peeking: if you look at the held-back result, change the model and test again, the held-back days have become fit days. Test once, at the end. The full routine is in how to backtest.

A made-up example: model against "tomorrow equals today"

Everything here is made up. A made-up asset starts at 100 and moves by the daily changes in the picture above. We use 24 daily changes. The first 12 are the fit period. The last 12 are held back.

The model is one line: tomorrow's change = k x today's change. The number k is chosen on the fit period only, by least squares. That gives k = 0.457. In words, the model says a rise is likely to be followed by another, smaller rise. The naive forecast says tomorrow's change is zero, which is the same as tomorrow's price equals today's.

Each forecast is scored by its miss: the absolute gap between the forecast change and the real change. The table gives the average miss per day.

On the fit period the model looks good: 0.833 against 1.036. A reader who stopped there would think it found something. On the held-back days the made-up market stopped moving in runs and began to zigzag, and the model's rule of "follow yesterday" was wrong again and again. Its average miss rose to 1.368, worse than the naive 0.942.

The lesson is not that models always fail. The lesson is that the fit period could not tell us. Only the held-back days could, and even then 12 days is a very small test. A real test needs many more observations and several different market conditions.

How to judge any prediction you read

When you see a bitcoin price prediction, ask these questions. A claim that cannot answer them is an opinion, whatever it looks like.

• Is there a number and a date? "It will go up" cannot be wrong. "It will close above X on date Y" can be. A claim that cannot be wrong cannot be scored.
• Is there a track record, and is it complete? A list of past calls that were right, with the wrong ones left out, tells you nothing. This is survivorship bias.
• Was it written before the event? A forecast posted after the move is a description.
• What is it compared with? A forecast is only useful if it does better than "tomorrow equals today" on data it did not see, after any costs.
• How many tries produced it? If someone tried hundreds of settings and showed the best, the result is partly luck.
• Who gains if you believe it? A seller of a course, a signal service or a coin has a reason to sound sure.

What a bot does and does not need from a forecast

People often assume a bot must predict the price to work. It does not. A bot is a set of rules that act on the price it sees now. The rules decide when to enter, how much to risk and when to stop. None of those steps needs a view of next month.

• It needs a clear rule. Written so exactly that a computer can follow it, and tested on data it has not seen.
• It needs a size. How much to put on each trade, so one bad run cannot end the account. See the position size calculator and risk of ruin.
• It needs a limit on loss. A stop, a cap on total exposure, a rule for when to switch off.
• It needs costs counted. Fees and the spread are paid whether or not the forecast was right.

What a bot does not need is a number for where the price will be. A grid bot, for example, places orders at fixed gaps and reacts to whatever the price does. If the price runs away in one direction, it can lose, and a forecast would not have fixed the rule. For how a bitcoin bot is put together, see the crypto trading bot guide.

If a bot does use a forecast, treat the forecast as one more part to test. Run it against the naive baseline, out of sample, with costs. If the bot only works when the forecast is right, you have built a bet on the forecast, and the sections above explain how little a fit to the past proves.

What this page does not tell you

It does not say that every model fails, and it does not say that testing makes a rule safe. A rule that passes a fair test can still lose money later, because the market can change. Testing removes some ways of fooling yourself. It does not remove risk.