def randwalk_drift(df, f_len, var):
#random walk with drift
#all f_base transformations
period = f_base.full_period(df, f_len, var)
df1 = f_base.hist_data(df, var)
df2 = f_base.randwalk(df, f_len, var)
slope = (df2.iat[-1, 0] + df2.iat[-1 * f_len, 0]) / (f_len - 1)
#calculates the slope of the line with respect to the points at
# final and initial business day
df2.loc[:, 'Forecast'] = df2.loc[:, 'Forecast'] + (f_len * slope)
#add slope times length to each element
df2.loc[:, 'Valid'] = df2.loc[:, 'Forecast']
#make new column called Valid which is the same as forecast
df = pd.concat([df1, df2], axis=1)
#combinde dataframes by Date, so there will be NaN values
# after today in the historica data value section
df = f_base.centralize(df, period[2])
return df
def savg(df, f_len, var):
#seasonal average (1-step)
#all f_base transformations
period = f_base.full_period(df, f_len, var)
df1 = f_base.hist_data(df, var)
df2 = f_base.lag(df, f_len, var)
df2.index = df2.index + pd.tseries.offsets.BDay(f_len)
#move forecast into the future so you don't have to make new columns
df3 = df2
#remember that they're both moved into the forecast future
df3 = df3.rename(columns={'Forecast': 'Valid'})
#create last dataframe for validation period
savg = df2.loc[period[3]:period[4], 'Forecast'].mean()
#calculate the mean of today back f_len business days
#doesn't seem right but the data was moved up, so it works
savg_v = df3.loc[period[1]:period[2], 'Valid'].mean()
#do same for validation
df2.loc[period[3]:period[4], 'Forecast'] = savg
df3.loc[period[1]:period[2], 'Valid'] = savg_v
#do same for validation period
df = pd.concat([df1, df2, df3], axis=1)
#put together all of the columns
df = f_base.centralize(df, period[2])
return df
def dses(df, f_len, var):
#Brown's Linear Exponential Smoothing/ Double exponential smoothing
#Trend is NOT removed
#all f_base transformations
period = f_base.full_period(df, f_len, var)
df1 = f_base.hist_data(df, var)
alpha = f_base.alpha(f_len)
smooth = overlay.ses_overlay(df, 1, var)
#smooths the dataframe
dsmooth = overlay.ses_overlay(smooth, 1, 'Overlay')
#double smoothed dataframe
dsmooth.index = dsmooth.index - pd.tseries.offsets.BDay(1)
#double smooth series is used after a single time period is removed
level = 2 * smooth - 1 * dsmooth
trend = smooth - 1 * dsmooth
df2 = level + (f_len * (alpha / (1 - alpha)) * trend)
period = f_base.full_period(df2, f_len, 'Overlay')
#take the period now because otherwise dates might not line up
df2 = df2.rename(columns={'Overlay': 'Forecast'})
#Rename becasue it's not an overlay anymore
df2.index = df2.index + pd.tseries.offsets.BDay(f_len)
#put it up to forecast because that's what the equation says
df3 = df2
#no need to move them because it's a continuous forecast#
df3 = df3.rename(columns={'Forecast': 'Valid'})
df = pd.concat([df1, df2, df3], axis=1)
#put together all of the columns
df = f_base.centralize(df, period[2])
return df
def ses(df, f_len, var):
#Brown's Simple Exponential Smoothing (1-step)
#all f_base transformations
period = f_base.full_period(df, f_len, var)
df1 = f_base.hist_data(df, var)
df2 = f_base.lag(df, f_len, var)
df2 = overlay.ses_overlay(df2, f_len, 'Forecast')
#gives you the ses and the last value is equal to the forecast
df2 = df2.rename(columns={'Overlay': 'Forecast'})
#rename the only column to forecast
df2.index = df2.index + pd.tseries.offsets.BDay(f_len)
#move up the data to forecast period
df3 = df2
#make validation column
df3 = df3.rename(columns={'Forecast': 'Valid'})
forecast = df2.iat[-1, 0]
#gives you the forecast value
preforecast = df2.iat[-2, 0]
#day before forecast
df2.loc[period[3]:period[4], 'Forecast'] = forecast
#make tomorrow till end = forecast value
df2.loc[period[2], 'Forecast'] = preforecast
#make today the preforecast value
#same for validatijon
valid = df3.at[period[3], 'Valid']
prevalid = df3.at[period[2], 'Valid']
df3.loc[period[1]:period[2], 'Valid'] = valid
df3.at[period[0], 'Valid'] = prevalid
df = pd.concat([df1, df2, df3], axis=1)
#put together all of the columns
df = f_base.centralize(df, period[2])
return df
def lin_reg(df, f_len, var):
#linear regression
'''
THIS DOES NOT WORK YET
'''
#all f_base transfromations
period = f_base.full_period(df, f_len, var)
df1 = f_base.hist_data(df, var)
df2 = overlay.ses_overlay(df, 1, var)
df2 = df2.rename(columns={'Overlay': 'Forecast'})
x = len(df2)
df2 = df2.loc[:, 'Forecast'].between(1, x + 1)
return df2
def randwalk(df, f_len, var):
#random walk
#all f_base transformations
period = f_base.full_period(df, f_len, var)
df1 = f_base.hist_data(df, var)
df2 = f_base.randwalk(df, f_len, var)
df2.loc[:, 'Valid'] = df2.loc[:, 'Forecast']
#make new column called Valid which is the same as forecast
df = pd.concat([df1, df2], axis=1)
#combinde dataframes by Date, so there will be NaN values
# after today in the historica data value section
df = f_base.centralize(df, period[2])
return df
def hles(df, f_len, var):
#Holts linear exponential smoothing
#Trend is NOT removed
#all f_base transfromations
period = f_base.full_period(df, f_len, var)
df1 = f_base.hist_data(df, var)
alpha = f_base.alpha(f_len)
beta = f_base.beta(f_len)
smooth = overlay.ses_overlay(df, 1, var)
#smooths the dataframe
dsmooth = overlay.ses_overlay(smooth, 1, 'Overlay')
#double smoothed dataframe
dsmooth.index = dsmooth.index - pd.tseries.offsets.BDay(1)
#double smooth series is used after a single time period is removed
#basic levels and trends
level = 2 * smooth - 1 * dsmooth
trend = smooth - 1 * dsmooth
#Make new data frames for the t-1 series
lesslevel = level
lesstrend = trend
#gives the level and trend at t - 1
lesslevel.index = lesslevel.index - pd.tseries.offsets.BDay(1)
lesstrend.index = lesstrend.index - pd.tseries.offsets.BDay(1)
#set original data to new dataframe with column named overlay
# so you can manipulate it with the others
original = df1
original = original.rename(columns={'Hist_Data': 'Overlay'})
#gives level and trend with respect to holts model
newlevel = alpha * original + (1 - alpha) * (lesslevel + lesstrend)
newtrend = beta * (level - lesslevel) + (1 - beta) * lesstrend
#there are a few gaps from holidays so fill those, but don't fill
# any possible future forecasts
newlevel = newlevel.bfill()
newtrend = newtrend.bfill()
forecast = newlevel + f_len * newtrend
#this is Holts final equation using the new trend and level
forecast = forecast.rename(columns={'Overlay': 'Forecast'})
forecast.index = forecast.index + pd.tseries.offsets.BDay(f_len)
validation = forecast
#validation column is the same as the forecast
validation = validation.rename(columns={'Forecast': 'Valid'})
#rename for error calculations
df = pd.concat([df1, forecast, validation], axis=1)
df = f_base.centralize(df, period[2])
return df
