Since the last upgrade to the Greed and Fear indicator, I have made new significant progress. As shown in the previous update, there was still some concern with the performance in the last couple of months of the simulation. After reconfiguring the neural network and re-transforming the input data, that concern is now gone. The performance now manages to go steadily higher in that last couple of months as well to end up somewhere around the 10000 points, much higher than previous results.
Of course, this improvement could be viewed as 'curve fitting' which is always difficult problem to avoid in building models. But here, that's not the case I think. In my previous post, I explained how the model is tested. In short it's by replaying history, and if done correctly, the model can of course not use future data as it goes through the historical data.
Now the latest improvement has been implemented, it doesn't mean a similar problem of failing performance will not happen again in the future. But it sounds reasonable to think that the more improvements are made, the less problems we will face moving forward. Time will tell.
Below is that chart with latest simulated results (click chart for bigger version, new page will open)
For more up to date developments, check out the results page.