An AI that wins every time in past markets is an AI that will lose in the future

“Almost undefeated in the past 5 years of testing.” When you see results like this, don’t buy out of hype—be wary. Extremely strong results are not a proof of strength but a red flag.

The reason is curve fitting, i.e., excessive tuning in Japanese. If you increase parameters and adjust repeatedly, you can create a “past-only” logic that clings to the details of historical charts. The more you cling to the past, the less adaptable it becomes to the future. Perfection in the past and functionality in the future become a trade-off at some point. An almost all-win test image is highly likely to show the extent of that trade-off.

To preemptively answer what distinguishes GENSEN-AI’s filtering from overfitting: overfitting optimizes for “past data” and boasts performance on the same data. Our filtering uses recent data that wasn’t used for training to generate results for selection. The separation between the data used for selection and the training data is the decisive difference.

Another point: among the top survivors in our validation, the win rate is 75.3% (settlements: 227 cases, validation figure). It is not a perfect all-win. 24.7% are losses. Please understand that this “presence of losses” is evidence that this is not an overfitted number.