Single point prediction is deceptive, because it uses the future data (test data) as the input and compares them to the prediction data and adjust the error rate inside. (?)
LSTM model is pretty similar to the random walk model, on the prediction data. (2.12)
Market capitalization (market cap) is the total market value of the shares outstanding of a publicly traded company; it is equal to the share price times the number of shares outstanding. Volume means the total number of shares/contracts that have been bought and sold of the particular company. (2.15)
It's weird to find that the amount of the bitcoin is not stable, even sometimes is decreasing.
It is said the total number of Bitcoin is 21,000,000. According to the Market Cap and the Close price, I calculate the total amount of the Bitcoin (I am not sure whether I use the right way). What I find now the highest amount of Bitcoin is around 20,346,312 and the mean value since 2018 is around 17,000,000, here I have to notice that time series of the bitcoin amount is not a simple increasing line as I thought, it is quite unstable. (2.20)
The random walk hypothesis is a financial theory stating that stock market prices evolve according to arandom walk (so price changes are random) and thus cannot be predicted. (2.20)
Notes for 02/12 Wed
(1) Differences between single point prediction and window based prediction are discussed. The reason of why single prediction is highly accurate is that previous prediction errors are not counted into current prediction. In other words, the coefficients in RNN for previous prediction are trained extremely small. (Personally, I think the RNN is trained to mimic the randomly walking model.)
(2) "History price vector based RNN" is proposed for window like prediction. However, the type, training data format and etc. require deep discussion and study.
(3) "Image matching" probably can be used for BitCoin price prediction. This thinking is based on the consideration that other relative mature markets can reflect people investment strategies which may cause the same trend in Bitcoin market. If so, CNN or RCNN may be applied
(4) In stochastic process analysis, "expectation" is usually utilized for revealing interesting characteristics. In this case, maybe how to use RNN making profit in BitCoin market is more attractive. However, it is noteworthy that from this point of view, the investment strategy design seems to become a key point which I think is too hard to finish it recently or even a economy background is necessary.
In conclusion, the next objective may focus on (2) and (3). The most important is how we predict BTP and how confident we have about the prediction results with the chosen model.
Notes for 03/18 Sunday night.
I have tried some new models to do the prediction, sadly, I got the same results, not meaning the same curve, but all of them were far away from the real value. I used trees, gradient booting and random forest, bagging, adaptive.. (all those model, I used the packages python already had.). Those results remind me of the prior work. Because in the code, I set a lag time, about 30 days, which I had to. However, bitcoins price changed dramatically, we used 30 days before to predict the future, the result showed how bad it was. (Also learning a new way, xgb, as a tool to deal with data and train the model, very quick..but its hard to say the correct rate. Still, I do not quite understand the concept of xgb).
I have also compared some old days bubble with the bitcoins price, maybe it can help me do the image matching part.
Notes for 03/30 Sunday night.
Dot-com bubble, somehow, is really like bitcoin price change trend. I show the comparison result in Details part. We can find if we enlarge the bitcoin price trend, both curve share similar trend, but bitcoin changes more dramatically (more dramatically than showing in the pics, because the timeline is different).
Notes for 04/02 Monday.
Mark the time point (2001, 2009) . Summarize what I have.
Notes for 04/09 Monday.
Make the plan for the next three weeks. I already used Autoregressive Model to predict the price. In time series analysis, ARIMA (autoregressive integrated moving average) is a very popular method. According to the Wiki, ARIMA is "fitted to time series data either to better understand the data or to predict future points in the series (forecasting) ". However, it is more difficult in mathematical model. I plan to use one week to understand the model, in both mathematical part and coding part, then use the other week writing python to realize it.