Разговор с ChatGPT об улучшении нейронной сети и торговле на Форекс - страница 8
![MQL5 - Язык торговых стратегий для клиентского терминала MetaTrader 5](https://c.mql5.com/i/registerlandings/logo-2.png)
Вы упускаете торговые возможности:
- Бесплатные приложения для трейдинга
- 8 000+ сигналов для копирования
- Экономические новости для анализа финансовых рынков
Регистрация
Вход
Вы принимаете политику сайта и условия использования
Если у вас нет учетной записи, зарегистрируйтесь
А все, рассмотрел
Переобученная сеть (подгон на истории) против правильно обученной сети. Это одна сеть, то есть решение бай или селл.
Прибыль на неизвестном периоде, который не входит в обучение.
Вывод: тысяча чертей, и так сойдет!!
Спрошу ChatGPT, что с этим делать...
Переобученная сеть (подгон на истории) против правильно обученной сети. Это одна сеть, то есть решение бай или селл.
Прибыль на неизвестном периоде, который не входит в обучение.
Вывод: тысяча чертей, и так сойдет!!
Спрошу ChatGPT, что с этим делать...
Это говорит о том, что любое обучение (пацанское или непацанское) дает случайные результаты на контролькой выборке, поскольку ты не можешь контролировать процесс обучения через ошибку на валидационной выборке, потому что все подвыборки несопоставимы.
Bear with me!- user: If you remember I am doing this for Forex, so the intentional overfitting shows a much worse loss on the unseen data, but it still shows a better profitability. While the prevention of overfitting shows a much better loss, but the profitability is worse. What do you think of this? Is it possible?
- robot: Considering that your model is intended for Forex trading, the primary objective is to maximize profitability on unseen data (live trading data) rather than just optimizing loss. Here are some insights regarding the situation you described:
1. **Profitability vs. Loss**: It is not uncommon for models in trading to exhibit a trade-off between loss metrics (e.g., MSE, cross-entropy) and actual profitability. In some cases, models that generalize well (prevent overfitting) may not perform optimally in real scenarios due to the inherent complexity and noise in financial markets.
2. **Risk-Return Trade-off**: The concept of intentional overfitting potentially capturing specific patterns in historical data to optimize for short-term profitability may lead to higher returns initially. However, this approach can be risky as the model might not generalize well and could fail in real-world trading scenarios or unseen data.
3. **Ensemble Models**: You could explore ensemble methods that combine models from both intentional overfitting and prevention of overfitting approaches to leverage the strengths of each method and potentially achieve a balance between profitability and generalization.
>>>>>>>> Я думаю, что я прав: оба варианта зарабатывают, значит, я все правильно делаю. А как зарабатывают, да хоть по Лунному календарю. Итог положительный, и это главное.
- user: If you remember I am doing this for Forex, so the intentional overfitting shows a much worse loss on the unseen data, but it still shows a better profitability. While the prevention of overfitting shows a much better loss, but the profitability is worse. What do you think of this? Is it possible?
- robot: Considering that your model is intended for Forex trading, the primary objective is to maximize profitability on unseen data (live trading data) rather than just optimizing loss. Here are some insights regarding the situation you described:
1. **Profitability vs. Loss**: It is not uncommon for models in trading to exhibit a trade-off between loss metrics (e.g., MSE, cross-entropy) and actual profitability. In some cases, models that generalize well (prevent overfitting) may not perform optimally in real scenarios due to the inherent complexity and noise in financial markets.
2. **Risk-Return Trade-off**: The concept of intentional overfitting potentially capturing specific patterns in historical data to optimize for short-term profitability may lead to higher returns initially. However, this approach can be risky as the model might not generalize well and could fail in real-world trading scenarios or unseen data.
3. **Ensemble Models**: You could explore ensemble methods that combine models from both intentional overfitting and prevention of overfitting approaches to leverage the strengths of each method and potentially achieve a balance between profitability and generalization.
>>>>>>>> Я думаю, что я прав: оба варианта зарабатывают, значит, я все правильно делаю. А как зарабатывают, да хоть по Лунному календарю. Итог положительный, и это главное.
Хорошо отвечает, это GPT4? или 3.5
Ты все делаешь правильно в своей ситуации, но это достаточно ненадежное обучение.
Хорошо отвечает, это GPT4? или 3.5
Стоит на планшете, в настройках указано Version 3.1 (0)
ого, даже так.
ого, даже так.
А это вариант для двух сетей: бай или неопределенность и селл или неопределенность.
По-прежнему, это прибыль на неизвестном периоде, который не входит в обучение.
И выходит, что переобучение все-таки не справляется. А это в свою очередь означает, что двухсетевой подход лучше чем односетевой. Или это означает, что результат по односетевому обучению был случайным.
Как советует ChatGPT, нужно все объединять...
А это вариант для двух сетей: бай или неопределенность и селл или неопределенность.
По-прежнему, это прибыль на неизвестном периоде, который не входит в обучение.
И выходит, что переобучение все-таки не справляется. А это в свою очередь означает, что двухсетевой подход лучше чем односетевой. Или это означает, что результат по односетевому обучению был случайным.
Как советует ChatGPT, нужно все объединять...
Тоже так думаю