Machine learning can identify patterns in financial markets without necessarily turning those patterns into money. A new study spanning 40 stocks, foreign exchange pairs and cryptocurrencies makes that distinction unusually clear: once transaction costs were included, only a small minority of the tested assets remained profitable.
The research, published on 1 October 2026 in Frontiers in Applied Mathematics and Statistics, evaluated decision trees, random forests and XGBoost models using one-minute market data. Rather than judging the algorithms only by whether they predicted price direction correctly, the researchers asked a harder practical question: could those predictions support profitable intraday trading after the friction of buying and selling was taken into account?
Why prediction and profit are different problems
Machine-learning studies of financial markets often report classification accuracy or related predictive measures. Those statistics matter, but a trading system faces an additional hurdle. A correct prediction is economically useful only when the price movement is large enough, frequent enough and captured efficiently enough to exceed transaction costs.
Alan Gaona, Jennifer Rangel-Madariaga and Lourdes Uribe therefore designed their analysis to separate predictive performance from economic performance. This is important because an algorithm can detect a statistically meaningful market pattern while still losing money when its signals are converted into trades.
The study covered 40 financial assets. The equity sample contained 24 stocks across sectors including information technology, consumer discretionary, consumer staples, communication services and financials. Ten foreign exchange pairs included major and more exotic currencies, while six cryptocurrencies included Bitcoin, Ether and several altcoins.
Two years of one-minute market data
The researchers analysed one-minute data from 1 January 2022 through 31 December 2023. For each asset, the dataset included open, high, low and closing prices and trading volume, together with widely used technical indicators intended to capture trends, momentum, volatility and market activity.
These indicators included simple and exponential moving averages, volume-weighted average price, moving average convergence divergence, the stochastic oscillator, momentum, rate of change, Bollinger Bands and the relative strength index. Fundamental information such as earnings, valuation ratios and macroeconomic indicators was deliberately excluded because the forecasting task operated over intervals measured in minutes rather than months or years.
Three tree-based machine-learning approaches were compared: a decision tree, random forest and XGBoost. The models attempted to predict sufficiently large upward or downward price movements at horizons of 1, 3, 5, 10, 15 and 30 minutes.
The analysis used walk-forward validation, meaning that models were repeatedly trained on historical information and then evaluated on later observations. Hyperparameters were tuned using Bayesian optimisation. After model selection, final configurations were evaluated across 30 predefined random seeds to assess algorithmic variability.
Crucially, the target itself incorporated trading costs. The researchers used standardised assumptions based on broker-reported spreads, with different cost levels for equities, foreign exchange and cryptocurrencies. Costs were then incorporated again when calculating the realised profit or loss from the simulated trades.
Longer horizons repeatedly came out ahead
The clearest pattern was not that one algorithm dominated every market. It was that the longer end of the tested forecasting range frequently performed better.
Among the 24 stocks, 22 selected the 30-minute forecasting horizon as their best configuration. All 10 foreign exchange pairs also selected 30 minutes. Four of the six cryptocurrencies favoured 30 minutes, while the other two selected 15 minutes.
This does not establish 30 minutes as the optimal horizon for machine-learning trading. It was the longest interval included in the study, so performance might continue changing at 45 minutes, an hour or longer. The result instead shows that within the six horizons actually tested, the models usually performed better toward the upper boundary. One possible explanation is that the shortest price movements contain too much market noise for the selected technical indicators and algorithms to extract stable directional information.
Model choice was more mixed. XGBoost was selected for 18 of the 40 assets, decision trees for 13 and random forests for nine. XGBoost was the most frequently selected model for stocks and foreign exchange, while random forest was selected most often among the six cryptocurrencies. This variation suggests that the most useful modelling approach depended on the asset rather than following a universal hierarchy.
Predictive performance looked better than trading performance
Stocks produced the strongest average predictive results among the three asset classes. Their average precision was 45.35%, average recall was 33.35% and the study’s precision-weighted F-beta score averaged 43.88%. Cryptocurrencies produced broadly competitive results, while foreign exchange pairs were more difficult for the models to predict.
Individual results also illustrate why a headline accuracy figure can be misleading. Apple’s selected decision-tree configuration, for example, achieved average precision of 51.70% but recall of only 16.75%. Tesla’s selected random-forest configuration reached 50.19% precision and 12.49% recall. The modelling objective deliberately placed greater weight on precision than recall, favouring more selective signals rather than attempting to identify every qualifying price movement.
The decisive test came when the researchers converted those signals into trades and deducted transaction costs.
Only five of the 40 assets, or 12.5%, generated a positive average return per trade. Six assets, or 15%, achieved positive cumulative profit over the two-year evaluation period. In other words, most of the models that found some predictive structure could not turn it into net economic gains under the study’s assumptions.
Foreign exchange produced the starkest warning
The contrast was especially clear in foreign exchange. The models generated measurable predictive performance, and every currency pair selected the 30-minute horizon, yet none of the 10 foreign exchange pairs generated positive economic returns after costs.
Cryptocurrencies performed better economically. Half of the six cryptocurrency assets generated positive cumulative profits. Among equities, three of the 24 stocks generated both positive average returns per trade and positive cumulative profit.
For the profitable stock configurations reported in the study, Tesla produced a cumulative profit of 1,463.08 under the simulation and an average return per trade of 0.0153%. Apple produced a cumulative profit of 145.39 and an average return per trade of 0.0045%, while Microsoft produced a cumulative profit of 463.40 and an average return per trade of 0.0020%. These figures are outputs of the study’s particular simulation and cost assumptions, not forecasts of future investment returns.
The researchers also evaluated profitable configurations using risk-adjusted measures including Sharpe, Sortino and Calmar ratios, as well as maximum tolerable slippage. Benchmark comparisons included passive, persistence-based and random-direction approaches. The profitable machine-learning strategies outperformed the random-direction benchmark, suggesting that their gains were not simply the result of arbitrary directional trades, although passive or simpler strategies could still outperform the models for particular assets.
What the findings mean for financial AI
The results reinforce a distinction that matters well beyond this particular set of algorithms. Financial prediction is not the same task as profitable execution.
A model can have genuine predictive information and still fail as a trading strategy because the predicted movements are too small to cover spreads and other costs. A model can also produce relatively precise signals too infrequently to generate meaningful returns. For practical financial AI, evaluation therefore needs to extend beyond conventional machine-learning metrics to include costs, risk, benchmarks and execution conditions.
The study also cautions against treating a particular algorithm as a universal winner. Although XGBoost was selected most often overall, the preferred algorithm, training-window size, prediction direction and forecasting horizon differed across assets. That heterogeneity is itself useful evidence. Financial markets are not one prediction problem repeated under different ticker symbols.
Important limits remain
The study is a historical modelling exercise rather than evidence that investors can reproduce these returns in live markets. Its one-minute dataset covered 2022 and 2023, so results may depend on the market regimes present during that period. Future market conditions can differ substantially.
The transaction costs were standardised within asset classes rather than reconstructed from the exact spread and execution environment surrounding every historical trade. Real-world trading also introduces latency, changing liquidity, market impact, taxes, brokerage arrangements and slippage that can differ across traders and platforms.
The analysis used technical and market-derived indicators and three tree-based model families. It therefore does not establish how alternative data, deep-learning systems or other forecasting methods would perform under the same framework. The 30-minute result also needs particular caution because 30 minutes was the longest horizon tested.
Even with those qualifications, the study provides a useful demonstration of why economic validation changes the story told by predictive statistics. Across 40 assets, machine learning frequently found its strongest signals at longer intraday horizons, but once realistic trading frictions entered the calculation, profitable opportunities became much rarer.
Source Information
Study: Intraday trading strategy selection under transaction costs with machine learning
Authors: Alan Gaona, Jennifer Rangel-Madariaga and Lourdes Uribe
Journal: Frontiers in Applied Mathematics and Statistics, Mathematical Finance
Published: 1 October 2026
DOI: 10.3389/fams.2026.1888767
