diff --git a/README.md b/README.md
index 5ded097..25b0014 100644
--- a/README.md
+++ b/README.md
@@ -14,6 +14,27 @@ suppress unnecessary components, enabling high performance in a wide range of
 scenarios. On a variety of real-world datasets, we demonstrate significant performance improvements over existing
 
 * [PyTorch Tutorial - Temporal Fusion Transformers](https://pytorch-forecasting.readthedocs.io/en/latest/tutorials/stallion.html)
+* [PyTorch Forecasting Announcement](https://towardsdatascience.com/introducing-pytorch-forecasting-64de99b9ef46#:~:text=What%20is%20PyTorch%20Forecasting%3F,easily%20trained%20with%20pandas%20dataframes.)
 * [PyTorch Tutorial - TimeSeriesDataSet](https://pytorch-forecasting.readthedocs.io/en/latest/api/pytorch_forecasting.data.timeseries.TimeSeriesDataSet.html#timeseriesdataset)
 * [Paper - Temporal Fusion Transformers
-for Interpretable Multi-horizon Time Series Forecasting](https://arxiv.org/pdf/1912.09363.pdf)
\ No newline at end of file
+for Interpretable Multi-horizon Time Series Forecasting](https://arxiv.org/pdf/1912.09363.pdf)
+* [Flight Demand Forecasting with
+Transformers](https://arxiv.org/pdf/2111.04471v1.pdf)
+
+
+* Check:
+  * UBER:
+    * [UBER BLOG](https://eng.uber.com/neural-networks/)
+    * [UBER PAPER](http://roseyu.com/time-series-workshop/submissions/TSW2017_paper_3.pdf)
+    * [UBER EXAMPLE](https://forecasters.org/wp-content/uploads/gravity_forms/7-c6dd08fee7f0065037affb5b74fec20a/2017/07/Laptev_Nikolay_ISF2017.pdf)
+    * [UBER EXAMPLE](https://conferences.oreilly.com/strata/strata-ca-2018/cdn.oreillystatic.com/en/assets/1/event/269/Detecting%20time%20series%20anomalies%20at%20Uber%20scale%20with%20recurrent%20neural%20networks%20Presentation.pdf)
+    * [UBER M4](https://eng.uber.com/m4-forecasting-competition/)
+  * [time-series-workshop](http://roseyu.com/time-series-workshop/)
+  * [Large-Scale Unusual Time Series Detection](https://robjhyndman.com/papers/icdm2015.pdf)
+  * [Mcompetitions code](https://github.com/Mcompetitions/M5-methods/tree/master/Code%20of%20Winning%20Methods)
+  * [Ocado Demand Forecasting](https://www.ocadogroup.com/technology/blog/how-we-improve-forecasting-and-availability-neural-networks-and-deep-learning)
+  * [Unit8](https://unit8.com/casestudies/)
+  * [TimeSeries_CNN_Classification_Using_Grid_Representation](https://github.com/junyoung-jamong/TimeSeries_CNN_Classification_Using_Grid_Representation)
+  
+  
+  
\ No newline at end of file
diff --git a/model/tft_regressor-1.pt b/model/tft_regressor-1.pt
new file mode 100644
index 0000000..142dd5c
Binary files /dev/null and b/model/tft_regressor-1.pt differ
diff --git a/predict_regressors.py b/predict_regressors.py
index e1c3db3..afefb72 100644
--- a/predict_regressors.py
+++ b/predict_regressors.py
@@ -17,12 +17,15 @@
 DROPOUT = spec["model"]["dropout"]
 HIDDEN_CONTINUOUS_SIZE = spec["model"]["hidden_continuous_size"]
 
-data = pd.read_csv("data/MERCHANT_NUMBER_OF_TRX.csv")
+data = pd.read_csv("data/poc.csv")
 data = data[[
     "MERCHANT_1_NUMBER_OF_TRX",
     "MERCHANT_2_NUMBER_OF_TRX",
-    "date"
+    # "USER_1_NUMBER_OF_TRX",
+    # "USER_2_NUMBER_OF_TRX",
+    "TIME"
 ]]
+data = data.rename(columns={'TIME': 'date'})
 data = data.set_index("date").stack().reset_index()
 data = data.rename(
     columns={
@@ -49,7 +52,8 @@
     .astype("category")
 
 # cut atypical values at the end of the sample
-train_data = data[:3200*2]
+cutoff = data["id"].nunique()
+train_data = data[:3200*cutoff]
 max_prediction_length = 24
 max_encoder_length = 72
 training_cutoff = data["time_idx"].max() - max_prediction_length
@@ -92,7 +96,12 @@
 
 model.load_state_dict(torch.load("model/tft_regressor.pt"))
 
-for target in ["MERCHANT_1_NUMBER_OF_TRX", "MERCHANT_2_NUMBER_OF_TRX"]:
+for target in [
+    "MERCHANT_1_NUMBER_OF_TRX",
+    "MERCHANT_2_NUMBER_OF_TRX",
+    # "USER_1_NUMBER_OF_TRX",
+    # "USER_2_NUMBER_OF_TRX",
+]:
 
     df = data[data["id"] == target].reset_index(drop=True)
 
diff --git a/train_regressors.py b/train_regressors.py
index 30c3434..fe0cbbd 100644
--- a/train_regressors.py
+++ b/train_regressors.py
@@ -61,7 +61,8 @@
     .astype("category")
 
 # cut atypical values at the end of the sample
-train_data = data[:3200*2]
+cutoff = data["id"].nunique()
+train_data = data[:3200*cutoff]
 max_prediction_length = 24
 max_encoder_length = 72
 training_cutoff = data["time_idx"].max() - max_prediction_length
diff --git a/train_regressors_benchmark.py b/train_regressors_benchmark.py
new file mode 100644
index 0000000..983c083
--- /dev/null
+++ b/train_regressors_benchmark.py
@@ -0,0 +1,69 @@
+import warnings
+
+import pandas as pd
+import numpy as np
+import matplotlib.pyplot as plt
+
+import statsmodels.api as sm
+# from statsmodels.tsa.arima_model import ARIMA
+
+warnings.filterwarnings("ignore")
+
+data = pd.read_csv("data/poc.csv")
+
+target = "MERCHANT_2_NUMBER_OF_TRX"
+data = data[[
+    target,
+    "TIME"
+]]
+data = data.rename(columns={'TIME': 'date'})
+# data["MERCHANT_1_NUMBER_OF_TRX"] = np.log(data["MERCHANT_1_NUMBER_OF_TRX"] + 1e-8)
+
+# add datetime variables
+data["month"] = pd.to_datetime(data.date).dt.month
+data["day_of_week"] = pd.to_datetime(data.date).dt.dayofweek
+data["hour"] = pd.to_datetime(data.date).dt.hour
+data["date"] = pd.to_datetime(data["date"])
+data = data.set_index("date")
+
+# cut atypical values at the end of the sample
+start = 3000 - 6*24
+train_data = data[:start]
+test_data = data[start + 1:3600]
+max_prediction_length = 24*15
+
+AR, I, MA = 5, 1, 5
+order = (AR, I, MA)
+
+SAR, SI, SMA, seas = 1, 0, 1, 24
+Sorder = (SAR, SI, SMA, seas)
+
+endog = train_data[target]
+
+exog = train_data.iloc[:, 1:]
+
+model = sm.tsa.statespace.SARIMAX(endog=endog,
+                                  exog=exog,
+                                  order=order,
+                                  seasonal_order=Sorder,
+                                  enforce_stationarity=False,
+                                  enforce_invertibility=False,
+                                  trend='c')
+
+res = model.fit(disp=True)
+
+exog_graph = None if exog is None else test_data[:max_prediction_length]
+
+fig, ax = plt.subplots(figsize=(7, 3))
+
+train_data[target][-200:].plot(ax=ax)
+test_data[target][:100].plot(ax=ax)
+
+# Construct the forecasts
+fcast = res.get_forecast(
+    steps=max_prediction_length,
+    exog=exog_graph.iloc[:, 1:]
+).summary_frame()
+fcast['mean'].plot(ax=ax, style='k--')
+# ax.fill_between(fcast.index, fcast['mean_ci_lower'], fcast['mean_ci_upper'], color='k', alpha=0.1)
+plt.show()
diff --git a/train_regressors_colab.py b/train_regressors_colab.py
index 789df5c..4022502 100644
--- a/train_regressors_colab.py
+++ b/train_regressors_colab.py
@@ -28,12 +28,15 @@
 HIDDEN_CONTINUOUS_SIZE = spec["model"]["hidden_continuous_size"]
 GRADIENT_CLIP_VAL = spec["model"]["gradient_clip_val"]
 
-data = pd.read_csv("/content/temporal-fusion-transformer/data/MERCHANT_NUMBER_OF_TRX.csv")
+data = pd.read_csv("/content/temporal-fusion-transformer/data/poc.csv")
 data = data[[
     "MERCHANT_1_NUMBER_OF_TRX",
     "MERCHANT_2_NUMBER_OF_TRX",
-    "date"
+    "USER_1_NUMBER_OF_TRX",
+    "USER_2_NUMBER_OF_TRX",
+    "TIME"
 ]]
+data = data.rename(columns={'TIME': 'date'})
 data = data.set_index("date").stack().reset_index()
 data = data.rename(
     columns={
@@ -60,7 +63,8 @@
     .astype("category")
 
 # cut atypical values at the end of the sample
-train_data = data[:3200*2]
+cutoff = data["id"].nunique()
+train_data = data[:3200*cutoff]
 max_prediction_length = 24
 max_encoder_length = 72
 training_cutoff = data["time_idx"].max() - max_prediction_length