How much data is needed to train a neural network?

Question

Neural networks

I tried to write a betting bot.
Used the FANN library .
Trained for 40,000 matches, when checking the next 10,000 unknown data, the network showed 66-69% correct results.
1. Is it possible to draw any conclusions from such a number of matches, or do you need a sample that is significantly larger than this one?
2. After each run of 40,000 data again, the results varied from 66 to 69%. Is this the norm or shouldn't it be? (The first thing that comes to mind is the random setting of the initial weights, and since the sample is not very large, it simply does not have enough data for training. This is where the error follows. Can you please comment on this situation?)
3. 66-69% for this amount of data, is this the norm or not?

Answer 1

1, 3. Difficult to analyze when you don't know anything about the application environment in which the network is used. Based on the input data of the network, the result is pretty good, it's not bad at all to win 66 times out of 100 (roughly speaking).
The neuron itself gradually determines which data has a greater influence on the result. But how can we know that exactly the data that we give as input is key? Or, perhaps, we did not submit all the data that is necessary for the input.
2. To increase the percentage of accuracy, you can play around with the learning rate, the number of intermediate layer nodes and epochs (but not the fact that this will help, perhaps this is the limit of the network).

Answer 2

It strongly depends on the completeness of the initial data, whether it is possible to build a pattern on them at all