1) Perhaps switch to conv1d instead of using conv2d layers in hyperparameter.py ? I'm pretty sure conv1d is just a layer on conv2d, but it's probably more reliable if things change in the future since i believe Tensorforce calls tensorflow's conv layers.
I'm working on a similar problem equities and this representation was giving me somewhat better results on the data. It gives the network two views of the data (detailed in the paper), which seems to help.
1) Perhaps switch to conv1d instead of using conv2d layers in hyperparameter.py ? I'm pretty sure conv1d is just a layer on conv2d, but it's probably more reliable if things change in the future since i believe Tensorforce calls tensorflow's conv layers.
2) Have you tried using this representation: https://arxiv.org/abs/1801.04503 for your data?
I'm working on a similar problem equities and this representation was giving me somewhat better results on the data. It gives the network two views of the data (detailed in the paper), which seems to help.