# TODO
# save / reload the model
#        self.serialize() => to_dict every where + json

# forecast future values => update datasets + forecast

# signal transformation => almost done
	- cumulative (integrated) tranformation
# Cutting into Estim/Valid (NE*H , NV*H) => cross-validation
	# ratio_E = 0.5
	# N_train = N_Estim + N_Valid + N_test (=H)
	# N_Estim = ratio_E * (N_train - H) 

# prediction intervals
	use cross validation 
	http://robjhyndman.com/hyndsight/tscvexample/
	book : https://www.otexts.org/fpp
	https://www.otexts.org/fpp/2/5

# improve plotting ... shaded area around prediction intervals
	neeed prediction intervals first

#Residual diagnostics	
	https://www.otexts.org/fpp/2/6

# other trends
	Exponential smoothing      ****************************************************
	exogenous variables   ********************************************
	moving average(N)
	moving median(N)

# other cycles
	seasonal
	user holidays etc (external tables?)

# other AR
  - ARX ,  VAR ?  *******************************************************
  - order control (look at timedelta ??).    ****************************************************

# Configuration (Options)
 - activate/disable transfromations/models/decomposition.  *******************************************
 - configure trends
 - configure cycles (CycleLength = ?)
	- cycle length should be in [5, 7, 12, 24 , 30, 60]
 - configure ARs (p = ?)
 - processing : threads etc
 - 

# Benchmarking
  - MComp
  - NN5
  - NN3
  - Yahoo stocks
	
# speedup things
  - python is sloooooooooooow (cython ?)
  - multiprocessing seems OK

#timedelta adaptive estimation
  - allow user control. *****************************************************
  - truncate timedelta to the nearest unit.  **************************************************
  - avoid saturday/sunday if not present in the dataset. *********************************************
  
# LOGGGING *********************************************************************

# cross validation for time series
 => http://robjhyndman.com/talks/MelbourneRUG.pdf

# feature selection (remove unnecessary lags and exogenous variables).
 ===> smaller model => smaller SQL code !!!


# real-life examples :
http://stackoverflow.com/questions/10302261/forecasting-time-series-data

# better graphics
https://stanford.edu/~mwaskom/software/seaborn/

# GitHub Topics
autoregressive benchmark cycles data-frame
exogenous forecasting heroku hierarchical horizon 
jupyter machine-learning-library
pandas restful-api scikit-learn seasonal
sql sql-generation time-series trends 

# Forecast Competition

http://eem2017.com/program/forecast-competition

In cooperation with our technical sponsor, we will provide you with a set of different weather input factors, e.g. wind direction, with which you are to forecast the power generation of a wind power plant portfolio. You may participate individually or as a team. The data input is organised in a realistic setting.