def model_ARIMA(key,train, test_shape, order, train_flag = 0, test = []):
predictions = []
rmse_val = []
if(train_flag):
test=test[0]
try:
train = train.values
except:
train = train
history = [np.asscalar(x) for x in train]
if train_flag==1:
itr=5
data=pd.DataFrame(history)
for i in range(3):
pred_temp = []
v_train=[np.asscalar(x) for x in data[:-itr].values]
v_expected=data.tail(itr).head(3).reset_index(drop = True)
try:
order=(order[0],1,order[2])
for j in range(3):
model = ARIMA(v_train, order = order)
model_fit = model.fit(disp=0)
yhat = model_fit.forecast()[0]
if yhat < 0:
yhat= mu.weighted_moving_average(history,1,3)
yhat=yhat[0]
pred_temp.append(yhat)
v_train.append(yhat)
except:
pred_temp.extend(mu.moving_average(v_train, 3 - len(pred_temp), 3))
mu.plotting(key, pred_temp, v_expected)
rmse_val.append(mu.calculate_rmse(key, v_expected, pred_temp))
if i == 2:
predictions.extend(pred_temp)
else:
predictions.append(pred_temp[0])
itr=itr-1
else:
try:
#TODO:check order
order=(order[0],1,order[2])
for t in range(test_shape):
model = ARIMA(history, order = order)
model_fit = model.fit(disp=0)
yhat = model_fit.forecast()[0]
if yhat < 0:
yhat= mu.weighted_moving_average(history,1,3)
yhat=yhat[0]
predictions.append(yhat)
history.append(yhat)
except:
predictions.extend(mu.moving_average(history, test_shape - len(predictions), 3))
predictions = [0 if pd.isnull(i) else int(i) for i in predictions]
return predictions,rmse_val
def model_LinearRegression(dataset=[], tsize=0, order=(0, 0, 0), train_flag=0):
predictions = []
rmse_val = []
if train_flag == 1:
itr = 5
for i in range(3):
expected = pd.DataFrame(dataset)
expected = expected.tail(itr).head(3).reset_index(drop=True)
train = dataset[:-itr]
diff_values = ut.difference(dataset, order[1])
scaler = ut.scaler_selection('lr')
diff_values = scaler.fit_transform(
pd.DataFrame(diff_values).values.reshape(-1, 1))
supervised = ut.timeseries_to_supervised(diff_values, order[0])
data = supervised.values
clf = LinearRegression()
param = {
"fit_intercept": [True, False],
"normalize": [False],
"copy_X": [True, False]
}
grid = GridSearchCV(clf, param, n_jobs=1)
model = mu.fit_model(data, grid)
for j in range(tsize):
X, y = data[:, 0:-1], data[:, -1]
yhat = mu.forecast_model(model, X)
# inverted = list()
# for i in range(len(yhat)):
# value = ut.inverse_difference(dataset, yhat[i], len(dataset) - i)
# inverted.append(value)
# inverted = np.array(inverted)
forecast = yhat[-1]
if forecast < 0:
forecast = mu.weighted_moving_average(dataset, 1, 3)[0]
predictions.append(forecast)
train = np.append(train, forecast)
diff_train = ut.difference(train, order[1])
diff_train = scaler.fit_transform(
pd.DataFrame(diff_train).values.reshape(-1, 1))
supervised = ut.timeseries_to_supervised(train, order[0])
data = supervised.values
predictions = predictions[1:4]
rmse_val.append(mu.calculate_rmse('GR_LR', expected, predictions))
itr = itr - 1
predictions = [int(i) for i in predictions]
return predictions, rmse_val
def model_DecisionTree(dataset=[], tsize=0, order=(0, 0, 0), train_flag=0):
predictions = []
for i in range(tsize):
diff_values = ut.difference(dataset, 1)
supervised = ut.timeseries_to_supervised(diff_values, 1)
data = supervised.values
if train_flag == 1:
train = data[0:-tsize]
else:
train = data
X, y = train[:, 0:-1].reshape(-1, 1), train[:, -1]
dtr = DecisionTreeRegressor()
param_tree = {
"max_depth": [3, None],
"min_samples_leaf": sp_randint(1, 11),
"criterion": ["mse"],
"splitter": ["best", "random"],
"max_features": ["auto", "sqrt", None]
}
gridDT = RandomizedSearchCV(dtr, param_tree, n_jobs=1, n_iter=100)
gridDT.fit(X, y)
clf = DecisionTreeRegressor(
criterion=gridDT.best_params_["criterion"],
splitter=gridDT.best_params_["splitter"],
max_features=gridDT.best_params_["max_features"],
max_depth=gridDT.best_params_["max_depth"],
min_samples_leaf=gridDT.best_params_["min_samples_leaf"])
clf.fit(X, y)
yhat = mu.forecast_model(clf, X)
inverted = list()
for i in range(len(yhat)):
value = ut.inverse_difference(dataset, yhat[i], len(dataset) - i)
inverted.append(value)
inverted = np.array(inverted)
forecast = inverted[-1]
if forecast < 0:
forecast = mu.weighted_moving_average(dataset, 1, 3)[0]
predictions.append(forecast)
dataset = np.append(dataset, forecast)
predictions = [int(i) for i in predictions]
return predictions
def model_ElasticNet(dataset=[], tsize=0, order=(0, 0, 0), train_flag=0):
predictions = []
for i in range(tsize):
diff_values = ut.difference(dataset, 1)
supervised = ut.timeseries_to_supervised(diff_values, 1)
data = supervised.values
if train_flag == 1:
train = data[0:-tsize]
else:
train = data
X, y = train[:, 0:-1].reshape(-1, 1), train[:, -1]
elas = ElasticNet()
param = {
"alpha": list(np.linspace(0.000000001, 100, 100000)),
"l1_ratio": list(np.linspace(0.000001, 100, 1000)),
"fit_intercept": [True, False],
"normalize": [True, False],
"precompute": [True, False]
}
random_elas = RandomizedSearchCV(elas, param, n_jobs=1, n_iter=100)
random_elas.fit(X, y)
clf = ElasticNet(
alpha=random_elas.best_params_["alpha"],
l1_ratio=random_elas.best_params_["l1_ratio"],
fit_intercept=random_elas.best_params_["fit_intercept"],
normalize=random_elas.best_params_["normalize"],
precompute=random_elas.best_params_["precompute"])
clf.fit(X, y)
yhat = mu.forecast_model(clf, X)
inverted = list()
for i in range(len(yhat)):
value = ut.inverse_difference(dataset, yhat[i], len(dataset) - i)
inverted.append(value)
inverted = np.array(inverted)
forecast = inverted[-1]
if forecast < 0:
forecast = mu.weighted_moving_average(dataset, 1, 3)[0]
predictions.append(forecast)
dataset = np.append(dataset, forecast)
predictions = [int(i) for i in predictions]
return predictions
def model_SVR_Poly(dataset=[], tsize=0, order=(0, 0, 0), train_flag=0):
predictions = []
for i in range(tsize):
diff_values = ut.difference(dataset, 1)
supervised = ut.timeseries_to_supervised(diff_values, 1)
data = supervised.values
if train_flag == 1:
train = data[0:-tsize]
else:
train = data
X, y = train[:, 0:-1].reshape(-1, 1), train[:, -1]
mod = SVR()
g = list(np.linspace(0.0001, 1, 1000))
C = list(np.linspace(0.01, 10, 25))
param = {
"kernel": ["poly"],
"degree": range(10, 30, 1),
"gamma": g,
"C": C
}
random_search = RandomizedSearchCV(mod, param, n_jobs=1, n_iter=100)
random_search.fit(X, y)
clf = SVR(kernel=random_search.best_params_["kernel"],
degree=random_search.best_params_["degree"],
gamma=random_search.best_params_["gamma"],
C=random_search.best_params_["C"])
clf.fit(X, y)
yhat = mu.forecast_model(clf, X)
inverted = list()
for i in range(len(yhat)):
value = ut.inverse_difference(dataset, yhat[i], len(dataset) - i)
inverted.append(value)
inverted = np.array(inverted)
forecast = inverted[-1]
if forecast < 0:
forecast = mu.weighted_moving_average(dataset, 1, 3)[0]
predictions.append(forecast)
dataset = np.append(dataset, forecast)
predictions = [int(i) for i in predictions]
return predictions
def model_RandomForest(dataset=[], tsize=0, order=(0, 0, 0), train_flag=0):
predictions = []
for i in range(tsize):
diff_values = ut.difference(dataset, 1)
supervised = ut.timeseries_to_supervised(diff_values, 1)
data = supervised.values
if train_flag == 1:
train = data[0:-tsize]
else:
train = data
X, y = train[:, 0:-1].reshape(-1, 1), train[:, -1]
rfr = RandomForestRegressor()
param_forest = {
"n_estimators": range(10, 1000, 100),
"criterion": ["mse"],
"bootstrap": [True, False],
"warm_start": [True, False]
}
gridRF = RandomizedSearchCV(rfr, param_forest, n_jobs=1, n_iter=100)
gridRF.fit(X, y)
yhat = mu.forecast_model(clf, X)
inverted = list()
for i in range(len(yhat)):
value = ut.inverse_difference(dataset, yhat[i], len(dataset) - i)
inverted.append(value)
inverted = np.array(inverted)
forecast = inverted[-1]
if forecast < 0:
forecast = mu.weighted_moving_average(dataset, 1, 3)[0]
predictions.append(forecast)
dataset = np.append(dataset, forecast)
predictions = [int(i) for i in predictions]
return predictions
def model_ARMA(key,train, test_shape, order, train_flag = 0, test = []):
predictions = []
rmse_val=[]
if(train_flag):
test=test[0]
try:
train = train.values
except:
train = train
history = [np.asscalar(x) for x in train]
if train_flag==1:
itr=5
data=pd.DataFrame(history)
for i in range(3):
pred_temp = []
v_train=[np.asscalar(x) for x in data[:-itr].values]
v_expected=data.tail(itr).head(3).reset_index(drop = True)
try:
for j in range(3):
model = ARMA(v_train, order = order)
model_fit = model.fit(disp=0, transparams=False, trend='nc')
yhat = model_fit.forecast()[0]
pred = yhat
if pred < 0:
pred = mu.weighted_moving_average(v_train, 1, 3)
pred = pred[0]
pred_temp.append(pred)
v_train.append(pred)
except:
pred_temp.extend(mu.moving_average(v_train, 3 - len(pred_temp), 3))
mu.plotting(key, pred_temp, v_expected)
if i == 2:
predictions.extend(pred_temp)
else:
predictions.append(pred_temp[0])
rmse_val.append(mu.calculate_rmse(key, v_expected, pred_temp))
itr=itr-1
else:
try:
for t in range(test_shape):
model = ARMA(history, order = order)
model_fit = model.fit(disp=0, transparams=False, trend='nc')
yhat = model_fit.forecast()[0]
inverted = list()
for i in range(len(yhat)):
value = ut.inverse_difference(history, yhat[i], len(history) - i)
inverted.append(value)
inverted = np.array(inverted)
pred = inverted[-1]
if pred < 0:
pred = mu.weighted_moving_average(history, 1, 3)
predictions.append(pred)
history.append(yhat)
except:
predictions.extend(mu.moving_average(history, test_shape - len(predictions), 3))
predictions = [int(i) for i in predictions]
return predictions,rmse_val
def model_ES(key, train, test_shape = 0, train_flag = 0, test = []):
predictions = []
rmse_val=[]
try:
train = train.values
except:
train = train
history = [np.asscalar(x) for x in train]
# TRAIN
if train_flag==1:
itr=5
data=pd.DataFrame(history)
for i in range(3):
pred_temp = []
v_train=[np.asscalar(x) for x in data[:-itr].values]
v_expected=data.tail(itr).head(3).reset_index(drop = True)
try:
for t in range(3):
if key=='SES':
model = SimpleExpSmoothing(history)
elif key=='HWES':
model = ExponentialSmoothing(history)
model_fit = model.fit()
yhat= model_fit.predict(len(history), len(history))
if yhat < 0:
yhat= mu.weighted_moving_average(history,1,3)
yhat=yhat[0]
pred_temp.append(yhat)
v_train.append(yhat)
except:
pred_temp.extend(mu.moving_average(v_train, 3 - len(pred_temp), 3))
mu.plotting(key, pred_temp, v_expected)
rmse_val.append(mu.calculate_rmse(key, v_expected, pred_temp))
if i == 2:
predictions.extend(pred_temp)
else:
predictions.append(pred_temp[0])
itr=itr-1
# FORECAST
else:
try:
for t in range(test_shape):
if key=='SES':
model = SimpleExpSmoothing(history)
elif key=='HWES':
model = ExponentialSmoothing(history)
model_fit = model.fit()
yhat= model_fit.predict(len(history), len(history))
if yhat < 0:
yhat= mu.weighted_moving_average(history,1,3)
yhat=yhat[0]
predictions.append(yhat)
history.append(yhat)
except:
predictions.extend(mu.moving_average(history, test_shape - len(predictions), 3))
predictions = [int(i) for i in predictions]
return predictions,rmse_val
def model_ML(dataset=[],
tsize=0,
test_shape=0,
model=np.nan,
key='',
order=(0, 0, 0),
train_flag=0):
predictions = []
pred_temp = []
rmse_val = []
parameter_values = {}
scale_flag = 0
if key == 'lr' or key == 'lasso' or key == 'ridge' or key == 'knn' or key == 'svmr':
scale_flag = 1
if train_flag == 1:
itr = 5
for i in range(3):
expected = pd.DataFrame(dataset)
expected = expected.tail(itr).head(3)
expected = expected.reset_index(drop=True)
train = dataset[:-itr]
diff_values = ut.difference(train, order[1])
if scale_flag == 1:
scaler = ut.scaler_selection(key)
diff_values = scaler.fit_transform(
pd.DataFrame(diff_values).values.reshape(-1, 1))
supervised = ut.timeseries_to_supervised(train, order[0])
data = supervised.values
RF_model = mu.fit_model(data, model)
pred_temp = []
for j in range(test_shape):
X, y = data[:, 0:-1], data[:, -1]
yhat = mu.forecast_model(RF_model, X)
#TODO: Inverse differencing and scaling
# if scale_flag==1:
# yhat=scaler.inverse_transform(pd.DataFrame(yhat).values.reshape(-1,1))
# if order[1]!=0:
# inverted = list()
# for i in range(len(yhat)):
# value = ut.inverse_difference(dataset, yhat[i], len(dataset) - i)
# inverted.append(value)
# inverted = np.array(inverted)
# forecast=inverted[-1]
# else:
# forecast = yhat[-1]
forecast = yhat[-1]
if forecast < 0:
forecast = mu.weighted_moving_average(dataset, 1, 3)[0]
pred_temp.append(forecast)
train = np.append(train, forecast)
diff_train = ut.difference(train, order[1])
if scale_flag == 1:
scaler = ut.scaler_selection(key)
diff_train = scaler.fit_transform(
pd.DataFrame(diff_train).values.reshape(-1, 1))
supervised = ut.timeseries_to_supervised(train, order[0])
data = supervised.values
pred_temp = pred_temp[1:4]
mu.plotting(key, pred_temp, expected)
if i == 2:
predictions.extend(pred_temp)
else:
predictions.append(pred_temp[0])
rmse_val.append(mu.calculate_rmse(key, expected, pred_temp))
itr = itr - 1
else:
dataset_1 = copy.deepcopy(dataset)
diff_values = ut.difference(dataset_1, order[1])
if scale_flag == 1:
scaler = ut.scaler_selection(key)
diff_values = scaler.fit_transform(
pd.DataFrame(diff_values).values.reshape(-1, 1))
supervised = ut.timeseries_to_supervised(diff_values, order[0])
data = supervised.values
RF_model = mu.fit_model(data, model)
try:
parameter_values = model.best_params_
except:
parameter_values = model.get_params()
test_shape = test_shape + 2
for i in range(test_shape):
X, y = data[:, 0:-1], data[:, -1]
yhat = mu.forecast_model(RF_model, X)
#
# if scale_flag==1:
# yhat=scaler.inverse_transform(pd.DataFrame(yhat).values.reshape(-1,1))
# if order[1]!=0:
# inverted = list()
# for i in range(len(yhat)):
# value = ut.inverse_difference(data, yhat[i], len(data) - i)
# inverted.append(value)
# inverted = np.array(inverted)
# forecast=inverted[-1]
# else:
# forecast = yhat[-1]
forecast = yhat[-1]
if forecast < 0:
forecast = mu.weighted_moving_average(data, 1, 3)[0]
predictions.append(forecast)
dataset_1 = np.append(dataset_1, forecast)
diff_values = ut.difference(dataset_1, order[1])
if scale_flag == 1:
scaler = ut.scaler_selection(key)
diff_values = scaler.fit_transform(
pd.DataFrame(diff_values).values.reshape(-1, 1))
supervised = ut.timeseries_to_supervised(diff_values, order[0])
data = supervised.values
predictions = predictions[2:test_shape]
predictions = [int(i) for i in predictions]
return predictions, rmse_val, parameter_values
def model_SVR_RBF(dataset=[], tsize=0, order=(0, 0, 0), train_flag=0):
predictions = []
for i in range(tsize):
diff_values = ut.difference(dataset, 1)
supervised = ut.timeseries_to_supervised(diff_values, 1)
data = supervised.values
if train_flag == 1:
train = data[0:-tsize]
else:
train = data
X, y = train[:, 0:-1].reshape(-1, 1), train[:, -1]
mod = SVR()
g = [
pow(2, -15),
pow(2, -14),
pow(2, -13),
pow(2, -12),
pow(2, -11),
pow(2, -10),
pow(2, -9),
pow(2, -8),
pow(2, -7),
pow(2, -6),
pow(2, -5),
pow(2, -4),
pow(2, -3),
pow(2, -2),
pow(2, -1),
pow(1, 0),
pow(2, 1),
pow(2, 2),
pow(2, 3)
]
C = [
pow(2, -5),
pow(2, -4),
pow(2, -3),
pow(2, -2),
pow(2, -1),
pow(1, 0),
pow(2, 1),
pow(2, 2),
pow(2, 3),
pow(2, 4),
pow(2, 5),
pow(2, 6),
pow(2, 7),
pow(2, 8),
pow(2, 9),
pow(2, 10),
pow(2, 11),
pow(2, 12),
pow(2, 13),
pow(2, 14),
pow(2, 15)
]
param = {'gamma': g, 'kernel': ['rbf'], 'C': C}
grid_search = RandomizedSearchCV(mod, param, n_jobs=1, n_iter=100)
grid_search.fit(X, y)
clf = SVR(gamma=grid_search.best_params_["gamma"],
kernel=grid_search.best_params_["kernel"],
C=grid_search.best_params_["C"])
clf.fit(X, y)
yhat = mu.forecast_model(clf, X)
inverted = list()
for i in range(len(yhat)):
value = ut.inverse_difference(dataset, yhat[i], len(dataset) - i)
inverted.append(value)
inverted = np.array(inverted)
forecast = inverted[-1]
if forecast < 0:
forecast = mu.weighted_moving_average(dataset, 1, 3)[0]
predictions.append(forecast)
dataset = np.append(dataset, forecast)
predictions = [int(i) for i in predictions]
return predictions
def training(details_data, datasets, forecast_period):
facc_out = dict()
rsq = dict()
price = details_data['price']
price = [float(i) for i in price]
price = pd.Series(price).fillna(0).tolist()
sku_list = details_data['sku']
market = details_data['market']
plant = details_data['plant']
spn = details_data['spn']
abc_data = details_data['abc_data']
#Profiling
data_prof = profiling.profiling_tech(datasets)
#Clustering based on nature
data_cluster = Cluster.clustering(data_prof)
total_price = np.sum(price)
sku_price = dict()
for i, sku in enumerate(sku_list):
sku_price[str(sku)] = price[i]
#Market Based Clustering
# plant_cluster=Cluster.clustering_plant(sku_list,plant,market)
#XYZ based on unit cost
xyz_data = xyz.xyz_class(sku_price, total_price)
#ABC based on volume
abc_alter = abc.abc_class(datasets)
# trained_outputs = []
forecast_results = []
num = 0
for incr, sku in enumerate(datasets):
num += 1
# if sku!='1702460700':
## if num!=1:
# continue
prof = data_prof.iloc[incr]
print("------------------------------------------------------------")
print("Running SKU %d: %s..." % (num, sku))
print("cluster : ", data_cluster[sku])
raw_data = copy.deepcopy(datasets[sku].T)
output = ut.init_output(forecast_period, raw_data, prof)
output['unit_cost'] = float(price[incr])
output['market'] = str(market[incr])
output['plant'] = str(plant[incr])
if pd.isnull(abc_data[incr]) == True:
output['Variability_Segment'] = abc_alter[sku]
else:
output['Variability_Segment'] = abc_data[incr]
output['Velocity_segment'] = xyz_data[sku]
output['spn'] = spn[incr]
dataset = raw_data.copy()
dataset = dataset[:-1]
# dataset = pp.dateformat(dataset)
# dataset, interval = pp.impute_missing_dates(dataset)
# print(interval.days)
if ((dataset['sales'] == 0).all() == True
or (set([math.isnan(x) for x in dataset['sales']]) == {True})):
# print(dataset['sales'])
print("All zeros/NaNs")
forecast = [0] * forecast_period
output['forecast_values'] = ut.assign_dates(
forecast, 'forecast', dataset.tail(1))
output['facc'], output['mape'], output[
'bias'] = ft.calculate_forecast_accuracy(
raw_data.iloc[-1], forecast[0])
facc_out[sku] = np.mean(
ft.calculate_validation_facc(forecast, forecast))
forecast_results = ft.output_forecast(sku, dataset,
datasets[sku].T, output,
forecast_results)
continue
sku_data = dataset.astype(np.float32)
sku_data = pp.read_from_first_sales(sku_data['sales'])
#size--->outlier bucket size
#sparse_size ---> number of zeros to categorize as sparse data
#freq ---> seasonality
interval = 30
size = 6
sparse_size = 10
freq = 12
# size,sparse_size,freq=pp.get_bucket_size(interval)
test_nan = pd.DataFrame(sku_data[-freq:])
test_nan = test_nan['sales']
#if last 1 year is NaN, impute data with zero and forecast is MA(6)
if sum(test_nan.isnull()) >= freq:
print("Last 1 year NaN")
sku_data = pp.data_imputation_zero(test_nan)
sku_data = sku_data[:-5]
expected = [0] * 5
forecast = mu.moving_average(sku_data, forecast_period, 6)
output['forecast_values'] = ut.assign_dates(
forecast, 'forecast', dataset.tail(1))
output['facc'], output['mape'], output[
'bias'] = ft.calculate_forecast_accuracy(
raw_data.iloc[-1], forecast[0])
facc_out[sku] = np.mean(
ft.calculate_validation_facc(expected, forecast))
forecast_results = ft.output_forecast(sku, dataset, sku_data,
output, forecast_results)
continue
#if # NaNs more than 60% impute with 0 else impute with values
if sum(pd.isnull(sku_data)) > (0.6 * len(sku_data)):
print("Nan Greater than 60%")
sku_data = pp.data_imputation_zero(sku_data)
else:
print("Nan less than 60%")
sku_data = pp.data_imputation(sku_data, freq)
sku_data = sku_data[0]
sku_data = pp.read_from_first_sales(sku_data)
#After reading from first non-zero if data is insufficient ---> weighted MA(3)
if len(sku_data) < 20:
try:
print("Weighted Moving Average")
forecast = mu.weighted_moving_average(sku_data,
forecast_period, 3)
output['forecast_values'] = ut.assign_dates(
forecast, 'forecast', dataset.tail(1))
output['facc'], output['mape'], output[
'bias'] = ft.calculate_forecast_accuracy(
raw_data.iloc[-1], forecast[0])
facc_out[sku] = ft.calculate_forecast_accuracy(
raw_data.iloc[-1], forecast[0])
forecast_results = ft.output_forecast(sku, dataset, sku_data,
output, forecast_results)
except:
print("Less than 3")
print(sku_data)
forecast = mu.moving_average(sku_data, forecast_period,
len(sku_data))
output['forecast_values'] = ut.assign_dates(
forecast, 'forecast', dataset.tail(1))
output['facc'], output['mape'], output[
'bias'] = ft.calculate_forecast_accuracy(
raw_data.iloc[-1], forecast[0])
facc_out[sku] = ft.calculate_forecast_accuracy(
raw_data.iloc[-1], forecast[0])
forecast_results = ft.output_forecast(sku, dataset, sku_data,
output, forecast_results)
continue
data_copy = sku_data.copy()
data_copy = np.array(data_copy)
# plt.figure()
# plt.plot(data_copy)
index1, index2, sflag1, sflag2 = pp.Sesonal_detection(sku_data)
sku_data = pp.outlier_treatment_tech(sku_data, interval, size)
sku_data = np.array(sku_data[0])
if sflag1 == 1:
sku_data[index1] = data_copy[index1]
if sflag2 == 1:
sku_data[index2] = data_copy[index2]
else:
sku_data = sku_data
# plt.plot(sku_data)
# plt.show()
# continue
sku_data = pd.DataFrame(sku_data)
#Testing Stationarity
d = 0
df_test_result = tests.dickeyfullertest(
sku_data.T.squeeze()) #pd.Series(sku_data[0])
while df_test_result == 0:
d += 1
if d == 1:
new_data = ut.difference(sku_data[0].tolist())
else:
new_data = ut.difference(new_data)
df_test_result = tests.dickeyfullertest(new_data)
sample = np.array(sku_data)
repeat = mu.check_repetition(sample, freq, 1, len(sample))
#Finding p and q value
try:
if d == 0:
p1, ps, pl = plots.acf_plot(sku_data, freq)
q = plots.pacf_plot(sku_data, freq)
data = sku_data
else:
p, ps, pl = plots.acf_plot(new_data, freq)
q = plots.pacf_plot(new_data, freq)
data = new_data
if repeat in ps:
p = repeat
elif repeat in pl:
p = repeat
else:
p = pl[0]
if p > freq:
p = freq
except:
p = 1
q = 1
data = sku_data
data = sku_data
best_order = (p, d, q)
print("BEST ORDER :", best_order)
#TODO: Calculate tsize
tsize = 5
# tsize = int(0.2*len(data))
# print(test)
expected = data[-tsize:].reset_index(drop=True)
expected = [float(i) for i in expected.values]
# print("Dimension: ", data.shape)
train_6wa = sku_data[0:-tsize]
predictions_ML, rmse_ML = train.time_series_using_ml(
sku_data, tsize, best_order, data_cluster[sku])
rmse_ARIMA, rmse_ES, rmse_naive, rmse_ma, predictions_ARIMA, predictions_ES, predictions_naive, predictions_ma = train.time_series_models(
freq, sku_data, data, tsize, best_order, data_cluster[sku])
print("Modeling done")
rmse_TS = rmse_ARIMA.copy()
rmse_TS.update(rmse_ES)
rmse_TS.update(rmse_naive)
rmse_TS.update(rmse_ma)
predictions = predictions_ML
predictions.update(predictions_ARIMA)
predictions.update(predictions_ES)
predictions.update(predictions_naive)
predictions.update(predictions_ma)
if data_cluster[sku] in [1, 4, 7, 10, 13, 16, 19, 22, 25]:
rmse_Croston, predictions_Croston = mu.Croston_TSB(sku_data, tsize)
rmse_TS.update(rmse_Croston)
predictions.update(predictions_Croston)
rmse_vol_ml = dict()
for key in rmse_ML:
std = np.std(rmse_ML[key])
mean = np.mean(rmse_ML[key])
rmse_vol_ml[key] = mean
# if std == 0:
# rmse_vol_ml[key]= mean
# else:
# rmse_vol_ml[key] = mean/std
rmse_vol_ts = dict()
for key in rmse_TS:
mean = np.mean(rmse_TS[key])
std = np.std(rmse_TS[key])
rmse_vol_ts[key] = mean
# if std == 0:
# rmse_vol_ts[key] = mean
# else:
# rmse_vol_ts[key]= mean/std
#Top 3 models
best_models_ml = sorted(rmse_vol_ml,
key=rmse_vol_ml.get,
reverse=False)[:3]
best_models_ts = sorted(rmse_vol_ts,
key=rmse_vol_ts.get,
reverse=False)[:3]
# forecasts_ml = dict()
# validation_ml = dict()
bias_ml = []
accuracy_ml = []
for model in best_models_ml:
# temp = ft.model_predict(model, best_order,data, forecast_period)
# forecasts_ml[model] = [0 if i < 0 else int(i) for i in temp]
# validation_ml[model] = predictions[model]
bias_ml.append(
(sum(expected) - sum(predictions[model])) / len(expected))
accuracy_ml.append(mu.calculate_facc(expected, predictions[model]))
bias_ml = [float(format(i, '.3f')) for i in bias_ml]
accuracy_ml = [float(format(i, '.3f')) for i in accuracy_ml]
# forecasts_ts = dict()
# validation_ts = dict()
bias_ts = []
accuracy_ts = []
for model in best_models_ts:
# temp = ft.model_predict(model, best_order, sku_data, forecast_period,repeat)
# forecasts_ts[model] = [0 if i < 0 else int(i) for i in temp]
# validation_ts[model] = predictions[model]
bias_ts.append(
(sum(expected) - sum(predictions[model])) / len(expected))
accuracy_ts.append(mu.calculate_facc(expected, predictions[model]))
bias_ts = [float(format(i, '.3f')) for i in bias_ts]
accuracy_ts = [float(format(i, '.3f')) for i in accuracy_ts]
#For one ensemble
error_ml = min(rmse_vol_ml.values())
error_ts = min(rmse_vol_ts.values())
best_models = [
min(rmse_vol_ml, key=lambda x: rmse_vol_ml.get(x)),
min(rmse_vol_ts, key=lambda x: rmse_vol_ts.get(x))
]
print("BEST MODELS :", best_models)
print("ERRORS OF BEST MODELS :", error_ml, error_ts)
forecast_ml, param_val_fore = ft.model_predict(best_models[0],
best_order, data,
forecast_period)
if best_models[1] == 'Croston':
rmse_Croston, forecast_ts = mu.Croston_TSB(sku_data,
forecast_period)
forecast_ts = forecast_ts['Croston']
else:
forecast_ts, param_val = ft.model_predict(best_models[1],
best_order, sku_data,
forecast_period, repeat)
forecast_ml = [0 if i < 0 else int(i) for i in forecast_ml]
forecast_ts = [0 if i < 0 else int(i) for i in forecast_ts]
weight_ts, weight_ml = ut.weight_calculation(data, best_models,
best_order)
print("weight ts:", weight_ts)
print("weight ml:", weight_ml)
Vm = predictions[best_models[0]]
Vt = predictions[best_models[1]]
Ve = ut.method_ensemble(Vm, Vt, weight_ml, weight_ts, tsize)
error_en = mu.calculate_rmse('Ensemble', expected, Ve)
bias_en = []
accuracy_en = []
bias_en.append((sum(expected) - sum(Ve)) / len(expected))
accuracy_en.append(mu.calculate_facc(expected, Ve))
bias_en = [float(format(i, '.3f')) for i in bias_en]
accuracy_en = [float(format(i, '.3f')) for i in accuracy_en]
#Ensemble of six month naive and weighted average
V6wa, rmse_6wa = ts.model_Naive('naive6wa',
train_6wa,
tsize, (0, 0, 0),
0,
train_flag=1)
error_6wa = np.mean(rmse_6wa)
forecast_6wa, param_val = ft.model_predict('naive6wa', best_order,
data, forecast_period)
forecast_en = ut.method_ensemble(forecast_ml, forecast_ts, weight_ml,
weight_ts, forecast_period)
output['forecast_period'] = forecast_period
output['interval'] = 'M'
output['best_models_ml'] = best_models_ml
output['best_models_ts'] = best_models_ts
output['bias_ml'] = bias_ml
output['bias_ts'] = bias_ts
output['bias_en'] = bias_en
output['accuracy_ml'] = accuracy_ml
output['accuracy_ts'] = accuracy_ts
output['accuracy_en'] = accuracy_en
output['TS'] = op.best_model_details_ts(best_models[1], bias_ts[0],
accuracy_ts[0], best_order)
output['ML'] = op.best_model_details_ml(best_models[0], bias_ml[0],
accuracy_ml[0], param_val_fore)
output['Ensemble'] = {"bias": bias_en[0], "accuracy": accuracy_en[0]}
error_min_model = min(error_ml, error_ts, error_en)
print("Errors:", )
print("ML:", error_ml)
print("TS:", error_ts)
print("Ensemble:", error_en)
print("six_naive_WA", error_6wa)
min_error = min(error_min_model, error_6wa)
if min_error == error_ml:
ftt = forecast_ml
elif min_error == error_ts:
ftt = forecast_ts
elif min_error == error_en:
ftt = forecast_en
else:
ftt = []
if min_error == error_6wa or all(elem == ftt[0]
for elem in ftt) == True:
print("Best forecast from six naive")
forecast = forecast_6wa
output['validation'] = ut.assign_dates(V6wa, 'validation',
dataset.tail(5))
validation_facc = ft.calculate_validation_facc(expected, V6wa)
output['validation_facc'] = ut.assign_dates(
validation_facc, 'val_facc', dataset.tail(5))
elif min_error == error_ml:
print("Best forecast from ML")
forecast = forecast_ml
output['validation'] = ut.assign_dates(Vm, 'validation',
dataset.tail(5))
validation_facc = ft.calculate_validation_facc(expected, Vm)
output['validation_facc'] = ut.assign_dates(
validation_facc, 'val_facc', dataset.tail(5))
elif min_error == error_en:
print("Best forecast from Ensemble")
forecast = forecast_en
output['validation'] = ut.assign_dates(Ve, 'validation',
dataset.tail(5))
validation_facc = ft.calculate_validation_facc(expected, Ve)
output['validation_facc'] = ut.assign_dates(
validation_facc, 'val_facc', dataset.tail(5))
elif min_error == error_ts:
print("Best forecast from TS")
forecast = forecast_ts
output['validation'] = ut.assign_dates(Vt, 'validation',
dataset.tail(5))
validation_facc = ft.calculate_validation_facc(expected, Vt)
output['validation_facc'] = ut.assign_dates(
validation_facc, 'val_facc', dataset.tail(5))
#
print("Forecasts:")
print("ML:", forecast_ml)
print("TS:", forecast_ts)
print("Ensemble:", forecast_en)
print("Best Forecast", forecast)
output['forecast_values'] = ut.assign_dates(forecast, 'forecast',
dataset.tail(1))
output['facc'], output['mape'], output[
'bias'] = ft.calculate_forecast_accuracy(raw_data.iloc[-1].sales,
forecast[0])
facc_out[sku] = np.mean(validation_facc)
output['forecast_ml'] = ut.assign_dates(forecast_ml, 'forecast',
dataset.tail(1))
output['forecast_ts'] = ut.assign_dates(forecast_ts, 'forecast',
dataset.tail(1))
output['forecast_en'] = ut.assign_dates(forecast_en, 'forecast',
dataset.tail(1))
output['model_ml'] = best_models[0]
output['model_ts'] = best_models[1]
forecast_results = ft.output_forecast(sku, dataset, sku_data, output,
forecast_results)
ft.plot_all_forecasts(dataset, sku_data, forecast, forecast_en,
forecast_ml, forecast_ts, sku)
return forecast_results, facc_out
