pic

PDESK FORCASTING

Price Prediction Product Launch

27th Feb, 2022

ABOUT NOBIS LAB

Nobis Labs is an AI-first forecasting platform for the commodities sector. We are currently working on the Palm, Soy complex for BMD, CBOT, Dalian, MCX and NCDEX exchanges. We build forecasting models for both Spot and Futures contracts.

WHY AI - A RULE GENERATION ENGINE

AI (Artificial Intelligence) or Machine Learning (ML) models use historical data (training data) to find patterns between inputs (features or x_variables) and outputs (y_variables or the variable to be forecasted). The model generates patterns (rules) between the x and y variables, without any human intervention. While a human can come up with 5 or 10 rules using say crop production data, technical features etc, the model can come up with 100`s of such rules. So the AI model can be thought of as a rule generation engine.

WHY AI - OBJECTIVE MODELS THAT HELP AVOID BIAS

AI models generate rules on historical data. They assign an importance to the features (inputs) using historical data. So, the models are very objective, leaving no room for bias.

Note: Lookahead refers to the number of future days for which the model is making a prediction. A 30 day lookahead implies that the model is trying to forecast 30 days into the future

HOW DOES THE PDESK PRICE PREDICTION PRODUCT HELP HEDGERS AND TRADERS

The PDesk price prediction product gives a 30 day lookahead (eventually 90 day lookahead) price prediction for the commodity of choice, every day. The prediction is updated daily on the cloud using fresh inputs (features). The list of features include the futures contracts of substitute commodities, forex prices, technical trading features, weather, trade and economic indicators data. The expansive list of features and the robust backtesting of models ensures that our model predictions are robust and driven by statistical relationships. The model outputs are consistent, systematic and will act as a reference/benchmark for your physical trading operations.

HOW DO OUR MODELS USE HUMAN INTELLIGENCE

We give the ability for users to input their insights into the model and use our What-If simulator. This lets the user use the model in a flexible fashion. If they don`t have any unique insights, they can look at the model`s base prediction. But if they do have some unique insights (let's say that they have some insights on the next month`s MPOB stock data), they can input their assumptions into the model and see how the modified prediction looks.

HOW DOES THE USER ACCESS THE PRODUCT

The user will access the product through a simple web dashboard.

MODEL PERFORMANCE BACKTESTS

A robust backtesting methodology ensures that the model`s performance can be reliably measured on historical data, without actually having to put the model into production and wait many years to measure model performance. The backtest is like a simulation, where the model is built using a dataset (training set) that is different from the one it is applied upon (testing set). This way, the model builder ensures that the model is not cheating and the backtest is a realistic assumption of historical performance.

IS MODEL PERFORMANCE ON THE BACKTEST A GUARANTEE OF FUTURE PERFORMANCE

Given the volatile and uncertain nature of financial markets, model performance in the future may be different from the past, even after robust backtesting. Past performance is not a guarantee of future performance.

HOW DOES MODEL BACKTEST PERFORMANCE LOOK FOR 3rd MONTH CPO(BMD) [3-CPO(BMD)]

Actual Price versus Model Predicted Price

The blue line is the actual price of 3-CPO(BMD), while the red line is the model predicted price of 3-CPO(BMD) [model prediction happened 30 days before].

Directional Correctness

Directional Correctness(%), or Accuracy(%), is the metric that tells us how correct the model was when it made binary decisions (price is expected to go up, price is expected to go down). The model`s numerical output (predicted price) is converted to a binary output and this metric is calculated.

Directional Correctness% for 3-CPO(BMD), 30-day lookahead

As can be observed from the above plot, the 30 day lookahead model crosses the 50% directional correctness threshold (red horizontal line) for 6 years, out of the 8 total years. For some years, the 30 day lookahead even touches the 70% mark.

Absolute Delta as a percentage of Actual Price

This metric calculates the absolute value of delta between predicted price and actual price and divides it by the actual price. It is expressed as a percentage.

Abs Delta as Percentage of Actual price(%) for 3-CPO(BMD), 30 day lookahead and all samples

The box plot should be interpreted as follows: the horizontal line inside each box is the median of the metric (for that year), the top of the box is the 75th percentile of the metric, the bottom of box is the 25th percentile of the metric and the extending lines/dots after the line denote outlier values in the metric.

It can be seen that the deltas have gone up for the year 2020, 2021, in line with the increased volatility due to covid related supply chain problems.

Would you like a 2 month free trial of the product? Contact us using this link

Stay tuned for more such blog posts