Structure of NN for input data with drop out

Question

Structure of NN for input data with drop out

Attack68

2020年8月14日 02:49

In financial markets, there is a simple problem of trading calendars varying across different countries. For example, Sweden observes Sweden National Day and Norway has Whit Monday. Typically, what happens then is that a time-series in the equity market that was closed for a holiday 'catches up' the next day when the market is open again.

For example:

+-----------------+------------+------------+------------+------------+
| Date            | SEK 1      | SEK 2      | NOK 1      | NOK 2      |
+-----------------+------------+------------+------------+------------+
| Date 1          | + 0.2%     | +0.4%      | +0.3%      | +0.6%      |
| Date 2          | + 1.1%     | +0.7%      | +0.3%      | +2.1%      |
| Date 3          | -3.2%      | -2.9%      | NaN        | NaN        |
| Date 4          | +0.1%      | -0.2%      | -2.8%      | -1.6%      |
+-----------------+------------+------------+------------+------------+

The objective of my model is to adjust the NOK returns for dates 3 and 4 since they have been distorted by the NOK holiday on Date 3. To do this I will use as many good dates as I have, such as date 1 and build a large dataset by randomly dropping out some data, adjusting the following the date and using the dropped out adjustments as the input and obviously I know the supervised output as the real data.

I feel that this would be well suited for a neural network, but I have never constructed one before with dropped out input.

Would it be appropriate to just build network where input data-points that are missing have no impact on the weights for that round? Are there any neural networks (or all) that do this by default?

Topic dropout neural-network

Category Data Science

hH1sG0n3 · Accepted Answer · 2020年8月14日 02:49

I do not think this bank holidays in your data should be treated as missing or nan. The markets are actually "not trading" on those days. This in itself is an intrinsic signal of the dataset which the model may favor to make better inference. For this reason, it should be explicitly set to zero.

Given big enough dataset and as far as neural networks go, LSTM models can serve you well in terms of capturing those drifts along many others you may not be able to pinpoint. This tutorial may be an excellent starting point of the pros and cons for your use case.

https://towardsdatascience.com/recurrent-neural-network-to-predict-multivariate-commodity-prices-8a8202afd853

Noah Weber · Accepted Answer · 2020年8月14日 02:48

There is no Special reason that neural networks are suitable, if anything it could be too much (depending on the complexities of your time-series).

In any case, you need to deal with you Nan values and if you are looking for a fast model, that imputes that automatically, I would suggest lgbm. Since it will impute NAN with values that reduce the out of fold loss the most (making the highly optimal imputation).

Structure of NN for input data with drop out

About