Forty teams entered; only six were still standing at the finish. A study published October 1 in *Frontiers in Applied Mathematics and Statistics* turned that brutal ratio into an academic conclusion: machine learning models found directional signals across all forty financial assets, but after deducting trading costs, only about 15% of assets remained cumulatively profitable.
The authors—Alan Gaona, Jennifer Rangel-Madariaga, and Lourdes Uribe—set themselves a far harsher task than prediction accuracy: do these predictions leave money on the table *after* friction?
🧪 How the experiment worked
The sample was forty assets: twenty-four stocks spanning five industry sectors from information technology to financials; ten forex pairs mixing major and emerging-market currencies; and six cryptocurrencies, led by Bitcoin and Ethereum plus several altcoins. All data were one-minute bars covering exactly two years, January 1, 2022 to December 31, 2023.
Features were nine families of textbook technical indicators:
- Trend and volume: simple and exponential moving averages, plus volume-weighted average price
- Momentum and oscillators: MACD and stochastic oscillator, momentum and rate of change
- Volatility and strength: Bollinger Bands and RSI
Three tree models were used: decision tree, random forest, and XGBoost. Prediction horizons came in six tiers, from 1 minute to 30, with 3, 5, 10, and 15 minutes in between. Validation used rolling walk-forward, hyperparameters were tuned with Bayesian optimization, and once a configuration was chosen, thirty random seeds tested stability.
The critical design choice was the objective function. Trading costs were set in normalized form based on broker-disclosed spreads, with separate fee schedules for stocks, forex, and crypto. Costs entered both the training objective and the final profit-and-loss accounting. The models knew from day one they were racing against the fee schedule.
📉 Predictions look decent; making money is another matter
First, the prediction side. Stocks averaged 45.35% precision, 33.35% recall, and a precision-weighted F-score of 43.88%. An intuitive reading of 45% precision: when the model calls a direction ten times, it's right fewer than five times. Crypto's forecasting performance was roughly on par with stocks; forex was hardest to predict.
Then the profit side: of forty assets, only about 15% stayed cumulatively profitable after costs. The gap between prediction and profit has a name: cost.
The table below shows the best-performing horizon automatically selected per asset class—the point is how consistent the pattern is:
| Category | Chose 30 min | Chose 15 min | Others | |---|---|---|---| | Stocks (24) | 22 | not individually disclosed | 2 | | Forex (10 pairs) | 10 | 0 | 0 | | Crypto (6) | 4 | 2 | 0 |
Model selection was far more scattered: XGBoost was chosen 18 times, decision tree 13, random forest 9; among cryptocurrencies, random forest was actually the most common pick. Horizon choice was highly consistent; model choice followed the asset. That itself is a conclusion: in minute-level trading, picking the right prediction horizon matters more than picking which model.
🤔 Three readings
First reading: short horizons die of noise. One explanation offered by the authors is that the shortest price moves contain too much noise for the chosen indicators and tree models to extract stable directional information. Much of the most active minute-level fluctuation is liquidity friction, not information.
Second reading: the 30-minute victory deserves an asterisk. It was the longest of the six candidate horizons; winning at the boundary doesn't mean there's nothing better beyond it. Whether 45 minutes or an hour would do better was not tested. Treating a boundary value as an optimum is the easiest way to misread studies like this one.
Third reading: the design is doubly strict on costs. Costs were deducted once in the training objective and again in the P&L accounting, and the thirty-seed stability check screened out single lucky runs. The 15% survival rate was computed under assumptions maximally favorable to the models—in other words, the real world would only be harsher.
🧭 Piecing it together with last week's paper
On September 30, zhichai.net published "The Price of Thought," whose conclusion was that adding reasoning compute to large language models does not buy net trading returns. This study used textbook tree models and technical indicators—no LLMs—yet the conclusions snap into the same puzzle piece: the signal exists, but cost and noise are the denominator, and the denominator is big enough to swallow the numerator.
For quant practitioners, the combined implication is quite concrete. The first paper cut down the superstition about inference budgets; this one cuts down the superstition about frequency: moving the prediction horizon from one minute to thirty minutes is more likely to change your P&L than swapping a decision tree for XGBoost.
One further limitation: the sample stops at the end of 2023. Those were high-volatility years; with a different period's cost structure and liquidity conditions, the survivor list might not be the same names.
---
Source: paper coverage in *Frontiers in Applied Mathematics and Statistics* (Research Today, 2026-10-01; authors Gaona, Rangel-Madariaga, Uribe).
Caveats: the original paper and DOI were not verified; all figures come from Research Today's retelling. Details such as the thirty random seeds and double cost accounting are likewise secondhand. Specific per-asset-class fee values were not given in the coverage. Data beyond 2024 were not covered.