Fixed the issue where tc_values df got values from both the gradient based learning as well as generative evolution

Final model. The model now has bootstrap based confidence testing, statistical testing and trust indicator scores. The calibration falls in the academically motivated thresholds and there is no Chelonsky decomposition failures in sight. Trying to get rid of the model outputs from the commit by clearing them and restarting the model

Final model. The model now has bootstrap based confidence testing, statistical testing and trust indicator scores. The calibration falls in the academically motivated thresholds and there is no Chelonsky decomposition failures in sight. Trying to get rid of the model outputs from the commit by clearing them and restarting the model

Final model. The model now has bootstrap based confidence testing, statistical testing and trust indicator scores. The calibration falls in the academically motivated thresholds and there is no Chelonsky decomposition failures in sight.

Added two-step calibration process in order to avoid local minima. This was done via differential evolution which made tc be pushed far enough out to be academically viable.

Added synthetic dataset creation. Omega initialization based on stationary dataseries to ensure proper initialization, added ADF tests to see how well dataset stationarised.

Added Lomb Scargle Spectral analysis on detrended time series in order to initialise omega value at the beginning of the training. Added the parameter constraints to be enforced during training and not after it. Added adaptive regularization to be based on exponential decay function in order to decrease bias while training the model. Next (and hopefully) last things to do are multi-window confidence testing as well as Bootstrap-Based trust indicator which assesses the robustness of the fit via resampling methods

Added the confidence calculation and visualisation for it. The new math in TF_boiler works now as well as does the training process with the hardcallback.

The iprovements made to the explicit matrix calculation model (now in Tensorflow model) taken from yesterdays build at test. The motivation behind changing the whole math process was due to the fact that in order to actually compute the confidence score utilising academically robust methods required a new look at the training process. Lets see how it goes

Model with lenghtening datasets. Next step the confidence indicator after which looking at explicitly restricting the values during the training process to be within the academically found threshold windows.

The shrinking window logic fixed and the model is expanding from the right direction. Next up the confidence score calculation.

The learning was from the end date onwards, which was wrong. Now the model is being thought from the earliest point to the latest. Now the shrinking window logic needs changing and the confidence score computation needs to be done as well.

Added threshold values and a hard cut system for fits which do not satisfy the threshold criteria proposed by Filimonov & Sornette 2013. Next step is the confidence score calculattion = qualified fits/all fits. After this implementation
''

Added L1 and L2 regularizations in to the code. Improved visualisation for the training and val loss plots

Added validation loss and cleaned out the code of the training loss visualisation. Early stopping is now checked against val loss and for 10 echelons. Next step is to update the visualisation so that all val and train losses are on the same grpah and to implement the hard/soft limit on parameters when constructing the final fit.

The model works now. Prints all the tc values in a single pd dataframe called res2.
Plots the LPPL estimates in a single graph with closing prices.
Plots histograms of frequency of estimate and a histogram of each fits predicted tc to detect if there is a pattern when the fitting window shrinks.

Made new plotting functions and got everything working.

The last code block produces a Dataframe which has all the lppl fits and closing prices information.
Next steps are to do the plotting functions now that the code is working properly. Also fixed the inverted prices problem.

Botoom code block plots out the shrinking time frame and cuts the start. Prior models cut the data from end date which was fixed.
Now the model inverts the time series which means that the effect is the same but can be fixed via either inverting the dataset or fixing the code.

value the mean of all values in the tc_values array.

Idea is that model files have the working basic and the test files contain new features that will be passed on to the main file once they are completely implemented.

Now the test file has implemented bayesien ideas by plotting histograms of TC dates that are calculated with shrinking time windows.
This allows us to examine the frequencies of each TC output and determine statistical measurements of the fits.

optimise multivariate regression model and plot
it better.

doing a multiple batch plot, confidence indicator
and other meaningful parameters.

Next steps: LPPL implementation to general linear model.

visualise the plot.

Converted the csv read to wti from brent
Still no idea why t1, t2, first and last parameters
are larger than tc which shouldn't be the case.

Current confidence parameter function not
working properly.

Datetime parsing fixed.

Visualisation works through notebook as wanted.

maplotlib inline, not working.

existing data provided. Trouble visualising.