Yes. We use different models for different time horizons and for different commodities. However, the models have a fairly similar high-level construction across the different horizons they work over.For a start, we incorporate measures of trends / seasonality in all our models, and these are very dependent on the time horizon over which we are forecasting. For example, in some commodities(e.g. industrial metals), there is strong evidence of trends over very short time horizons, but relatively little seasonality. And in others (e.g. energies), we might have seasonality exhibited over a year, but very little trending behaviour.Furthermore, our usage of alternative data is predicated heavily on certain time horizons. For example satellite data above open cast mines being particularly useful in making predictions over horizons between 1 month and 3 months into the future.
What kind of AI techniques are you using?
We are using a variety of machine learning models at different points in our prediction processes, for example LSTMs, Gaussian Processes, Support Vector Machines, LASSO and Multi-Task Learning.Aside from prediction accuracy, we also believe it is very important for our predictions to be explainable, hence our choice of techniques.
Our risk / compliance department is concerned about model transparency — How do you ensure your models can be explained?
We are motivated by explainability and also the ability to control our sensitivity to various inputs so that users can up/down weight the importance of economic data, currency movements, demand/supply information etc.
What commodities are you currently able to produce predictions for?
Aluminium, Copper, Nickel, Zinc, Steel Rebar, Iron Ore, Lead and Brent
How do you measure accuracy?
There are many different ways of looking at accuracy, we use the following 12 different measurements, clients are generally interested in directional and 1-MAPE:
- Percentage Directional Accuracy
- 1-MAPE (price)
- 1-SMAPE (price)
- Brier Score (returns)
- MAD from mean (price)
- MAD from median (price)
- Rank and Linear Correlation (returns)
- Prediction Standard Deviation (price)
- Prediction Standard Deviation (returns)
- Implied Volatility to Realised Volatility Ratio (returns)
- Percentage of Actual Inside Bound
Can you expand on how the AI is used and how it’s a differentiator?
At a very high level it’s used to integrate a variety of data sets that are pretty different to each other in nature (text, images, freight movement, economic time series) and sampled at different frequencies (daily, weekly, monthly etc), and in some cases not even consistently themselves (for example satellite imagery because of cloud cover). Furthermore, the data often has a very low signal to noise ratio and has a prognostic ability that changes with time, so that relationships between our inputs and what we are trying to predict changes.
How much does your service cost?
Our pricing is based on the number of commodities taken, and the size of the organisation. We employ a subscription model with a fee for each commodity per month that is tailored to your organization’s use case and number of users.
Could we have predicted the coronavirus?
No. However, as it’s an event that gets rapidly factored into the markets and all the data sources we monitor, we are well placed to make predictions relating to prices pretty soon after it all started. Please have a read of our recent blog on uncertainty.
What’s actually funky about the AI we use?
We have been applying state of the art techniques from the machine learning academic community for over a decade to the problem of predicting commodity prices. These are focussed on many different areas, for example:
- Ascertaining which candidate input sources have useful information content for the predictions we wish to make.
- Taking these input sources that are sampled at different frequencies (and not always consistently for example satellite data) and are very different to each other and highly noisy and extracting the signals they contain.
- Integrating these signals in a way to give consistent predictions along with confidence boundaries that describe our confidence in these predictions, over different time horizons.
- Taking advantage of the relationships between the commodities we are trying to predict themselves.
Do we just share raw predictions or do we also share reasoning / methodology?
We believe AI can only be truly useful if people do not treat it like a magic, black box and it is demystified so that business users can place their faith in it. This is particularly important in environments that are highly regulated. We have therefore baked explainability into every aspect of our prediction process. A result of this is that we are able to accompany every prediction with explainability weightings that show how the specific prediction was derived from the eight different families of data that feed into our models.
ChAI is helping multiple organisations along the supply chain in managing their exposure to raw material price risk through forecasting the price of raw materials. Please get in touch if you are interested in finding out more.
Originally published at https://www.chai-uk.com on April 14, 2020.