Choosing between algorithm families
When people say they used machine learning to forecast a cryptocurrency, they are usually naming an algorithm family: a recipe for turning examples into a function that maps inputs to an output. The families differ in how flexible they are, how much data they need, and how easy they are to interpret. This article compares the common ones and their typical weaknesses. It gives no forecasts, and no family is shown here to work reliably. It is not financial advice.
A theme runs through the comparison: more flexible models can fit more patterns, including patterns that are only noise. With the small, noisy, changing data typical of crypto markets, flexibility is often a liability.
Linear and regularised regression
Linear regression fits a weighted sum of inputs to the target. Variants such as ridge and lasso add a penalty that keeps weights small or sets some to zero, which reduces overfitting.
Strengths: it is fast, transparent, and each weight can be inspected. It also makes a strong baseline, and if a complex model cannot beat it, the complexity is not worth the trouble.
Weaknesses: it can only capture straight-line relationships unless features are engineered, and it assumes the relationships are stable over time. When price changes carry little predictable signal, fitted weights are typically close to zero, so its forecasts hover around “no change.”
Logistic regression and simple classifiers
For an up-or-down question, logistic regression outputs a probability of “up.” Its strengths and weaknesses mirror linear regression. A further caution applies: if the data has more up days than down days, a classifier can reach high accuracy by always saying “up.” Accuracy alone says little, and the article on evaluating predictions explains why.
Decision trees, random forests and boosting
A decision tree splits the data with a series of yes or no questions, such as whether recent volatility is above a threshold. Random forests average many trees trained on different samples, and gradient boosting builds trees one after another, each correcting the errors of the earlier ones.
Strengths: they handle non-linear effects and interactions between inputs without much manual work, cope with mixed data types, and often perform well on tabular data in many fields.
Weaknesses: single trees overfit easily. Boosted models can chase noise if not tuned carefully, and tuning on the same data repeatedly causes hidden overfitting. Trees cannot extrapolate: they can only output values within the range they have seen, which is awkward for a market that sets new highs or lows. They also have no built-in sense of time order unless lagged features are supplied.
Support vector machines and nearest neighbours
Support vector machines find a boundary separating classes, and nearest-neighbour methods predict from the most similar past cases. They were popular in earlier research on financial data. Their main weaknesses are sensitivity to how features are scaled, to the choice of settings, and, for nearest neighbours, to the assumption that situations that look alike in the past will lead to similar results. In markets where conditions rarely repeat, “similar” is hard to define.
Recurrent and sequence networks
Recurrent networks, including LSTM networks, are designed for sequences and keep a memory of earlier steps. Newer sequence models such as transformers handle long inputs differently. They are common in papers on crypto prediction because prices form a sequence.
Strengths: they can model complex dependencies over time and combine many input types, including text.
Weaknesses: they need large amounts of data, which crypto history does not supply for independent market cycles. They have many settings, are hard to interpret, and can overfit dramatically. A frequent finding is that such a network produces a line of predictions that closely follows the actual price but lags it by one step, essentially copying the latest value. That looks impressive on a chart while adding no real foresight.
Feed-forward neural networks
Ordinary multi-layer networks take a fixed set of inputs and output a prediction. They share the strengths and weaknesses of recurrent networks in lighter form: flexibility, data hunger and difficulty in understanding why they produce an output. Their results can also vary from run to run because of random starting values, so a single run reported in an article may not be representative.
Ensembles and hybrids
Many studies combine methods, for example a statistical volatility model with a neural network, or average several algorithms. Ensembles can reduce variance, which helps against overfitting. They do not remove the basic problem that the signal in the data is weak, and they add layers that are harder to audit.
Weaknesses that apply to all of them
- Weak signal: if price changes are close to a random walk, no algorithm can extract much from past prices alone.
- Non-stationarity: relationships change as markets, products and rules change.
- Limited independent data: a few cycles, however finely sliced, are still a few cycles.
- Evaluation errors: leakage, shuffled splits and repeated tuning inflate reported results, as detailed in the guide to model training.
- Costs: fees and the bid-ask spread can erase a small edge.
A practical way to compare
If you read a study or product claim, check which family was used, what baseline it was compared with, whether testing was done in time order on unseen data, and whether costs were included. Often a simple model with honest testing tells you more than a complicated one with a flattering chart. Cryptocurrency prices are volatile, past performance does not predict future results, and people can lose everything they invest.