3 ms·
> I should have taken the hyperparameters that yielded the best model for the latest evaluation set. > I have done enough research on production ML now to know
by wodenokoto 4y ago
> I should have taken the hyperparameters that yielded the best model for the latest evaluation set.
> I have done enough research on production ML now to know that it pays to simply overfit to the most recent data and constantly retrain. Successful companies do this.
I don’t understand why that is
- fxtentacle 4y agoBecause most old data becomes useless as the overall market situation changes. Real estate trends from 2019 won't help you predict much in 2022.
- spywaregorilla 4y agoThat's an utterly nonsense claim. The time sensitivity of data is highly context dependent and even if it's true that the real estate trends of 2019 won't help you predict 2022, it's crazy to assume that "overfitting" to 2021 will do better. I mean even this exact example. You think the patterns of 2021 peak covid real estate are a good example to learn from for 2022? No, absolutely not.