1.A Hybrid Model Approach for Price Discovery System

After the pre-processing steps, it is obvious that crop specific time-series analysis   is not possible due to lack of available data points. To overcome this problem, we have considered the following two sets of data for model building:

1. SET 1: Day wise average price of crops (time series data)
2. SET 2: Deviation of crop price across the AMCs (considering the specific crop type, variety etc.) from the day wise average price.

Now, we feed the SET 1 data to train a sequential model, which will give us a price prediction irrespective of AMC, crop type, variety etc. We called it as Global Price. Crop price may vary due to various reasons, such as inflation, high or low yield, weather condition etc. To identify such influencing factors, we required a considerable amount of data which is absent in present scenario. To tackle such conditions, we introduce the concept of Global Price. 

On the other hand, SET 2 data fed to a regression model which will give us the variation of crop price from the Global Price across the different AMCs (considering the specific crop type, variety etc.) with in the state. The output of this model is indicating the fluctuation of prices from the Global average prices among the AMCs. We call this predicted value as Local Price Deviation.

By combining the outputs from the said two models, we can predict the final crop price. Thus, our final prediction is a result of combining the output from the sequential model and the regressor. So, we termed our model as a hybrid model for price prediction.

Predicted Price = Global Price (Sequential model output) + Local Variation (Regression Model Output)

2. Approach for Volume Managment System

After observing the distribution of the data, we will build the following regression models:

Regression Model: This model will be used to predict Occupancy depending on input variables.

2.1 Train-Test Split of the Processed Data

We are splitting 80% as training data and 20% as testing data for building and testing the model respectively.

Table: Train-Test Data

2.2   Model Development

A number of models have been explored to fit the volume data and predict occupancy. Based on the performance of the different models, finally XGBoost model has been selected.