My capstone project at Galvanize, I aimed to predict a range of potential corn prices using data on the corn market and the weather in key producing states. I was able to achieve sufficient results for implementing a hedging strategy, although overall returns would not be high enough to warrant speculation. For current predictions and a prettier presentation, please visit my website: cooperscorn.com

Going into the project, I had two main goals:

• Predict the direction of the price movement to lend farmers confidence for hedging their crop production
• Develop a model that would earn acceptable investment returns for traders

To achieve these objectives, I developed several neural networks, which relied on market data as inputs. While I largely accomplished my goal of predicting direction, investment returns would be substandard for all but those with a very low cost of capital.

I evaluated my models using RMSE, hit rate (the percentage of times I predicted the correct direction of price movement), and trading return. The graph below shows my 6-month prediction model for both in-sample and out-of-sample results. The shaded regions show a confidence interval, as represented by 2 x RMSE.

![Image of Results] (Figures/neuralnet_six_month.png)

I am most pleased with my hit rate results, correctly identifying the direction of price movement 70% of the time when the current price is at least 10% different from my predicted price.

![Image of Results] (Figures/profits_hist_six.png)

Through the EDA process, I also uncovered some interesting phenomena in the data. As shown below, the introduction of corn syrup and E85 produced demand shocks that appear to be correlated with a steep increase in price. One can immediately see the dynamics of supply and demand over time in the second graph, as the increase in production has led to a consistent decline in inflation adjusted price.

![Image of Demand Shocks] (Figures/inflation_adjusted_demand_shocks.png)

![Image of Demand Shocks] (Figures/price_vs_production.png)

• Python Packages

  • Keras
  • Pandas
  • Numpy
  • SciPy
  • StatsModels
  • IbPy
  • Flask

• AWS EC2

While not all datasets collected were used as features for the final models, the following was collected.

• Corn Futures Prices
• Soybean Futures Prices
• Oil Futures Prices
• $USD Exchange Rate Index

• Corn Supply Levels
• Acreage Planted
• Yield Estimates

• Precipitation and Temperatures for Iowa, Illinois, Nebraska, Minnesota, and Indiana
• ONI Index

The final models I developed are feed-forward neural networks, which use either 6-month or 3-month lagged data as features. The structure of the nets is quite simple, consisting of 3 layers, the hidden layer having a linear activation function and the output layer having a sigmoid activation function. I used the lagged features rather than current data to simulate forecasting 3 or 6 months out with the information that is readily available.

After finishing the modeling phase of the project, I developed a website, cooperscorn.com, to display the results of my research as well as current predictions. The website was built using flask bootstrapping and is hosted on an Amazon EC2 instance.

• Monitor my predictions to ensure outcomes are comparable with current out-of-sample forecasts
• Speak with futures traders to identify other potentially import features
• Develop models for other crops