April 25, 2024
You do not need me to say that AI is a big deal these days, with Gen AI and LLMs are getting most of the attention. But I will make you notice that AI role in time-series is being ignored by the mainstream.
And that's good. This is the right environment to be able to create new things under the radar.
We are focusing on new RNNs developments at Pantar.ai, but let me tell you something first. Did you know for example that NASA uses time-series forecasting while observing solar winds on a minute-by-minute basis?
By the way, what is a time-series?
Price of coffee:
Time-series is what stock and bond markets are made of: a specific asset value per second, minute, day - a source of valuable information.
Historically, ARIMA, SARIMA and other statistical models have been used successfully to do forecasting using single variables: "dear model, take the average air daily temperature of one day in a given coastal town and forecast what the ocean temperature will be on that given day"
But those solid models tend to miss out things when we add multiple variables:"dear model, take the average air daily temperature of one day, the sun angle, the ocean currents, el niño, local pollution levels, etc in a given coastal town and forecast what the ocean temperature will be on that given day". This where AI deep learning models over-perform ARIMA (and others) in multivariate time-series regressions.
Let's be mindful that AI can help but does not really know the future. So you will hear many people who will claim that they are experimenting AI (time-series applications) to find the holy grail in investing, i.e. the way to always over-perform markets. Bulls**t alert. Any AI model over-fitting the past (= beating past stock market performance by much) will fail to properly predict the future, and thus over-perform in the future.
Why?
I say that in markets there is too much "white noise": one day assets are moved by inflation, the other by geopolitical events, then by lack of liquidity without established patterns and connections. There are further reasons here, for example, on why precisely forecasting markets is currently not possible.
But we can still do some good stuff with AI in markets. We are in fact experimenting with recurring neural networks (RNNs) to understand markets in real time.
(pic credit to Trist'n Joseph - article here)
RNNs are networks with feedback loops that allows information to persist (=like human memory) - to help us not to forecast (it is hopeless in markets) but to understand in real time the possible relationships between millions of variables and the price of an asset TODAY.
Mixing human trading experience, historical correlations and AI understanding of today (without forecasting tomorrow) can greatly enhance our ability to adapt in real-time our investment portfolios to evolving market conditions.
