So said Dr Ernest Chang on yesterday’s Zoom presentation at the STA’s monthly meeting. Known as Ernie among friends, he is charming, witty and speaks perfect English – albeit with a Chinese accent. With my very limited knowledge of all the sciences, I had been worried about covering this for our readers. I need to have done so: his presentation was clear, to the point, and avoided unnecessary jargon.
Introduced by Jeff Boccaccio, who also moderated the questions, both gents did very well with this new format we are all having to get used to. The doctor started at IBM’s famous Watson laboratory where he specialised in Machine learning – hence the title of his talk. He then moved into banking at Morgan Stanley and Credit Suisse; now he’s a Commodity Trading Advisor at QTS Capital Management LLC with 3 books under his belt: Quant Trading, Algo Trading and Machine Trading.
For his presentation he chose to focus on just one of his trading strategies, one based on day-trading e-mini S&P 500 futures – he does not trade options as he believes in simple strategies. The abundance of short term intra-day data is also useful for training purposes. Therefore, he also believes picking up one of the myriad existing black boxes (readily available for Python) and tweaking it via machine learning – which are not market specific.
He acknowledges, and warns, that humans are often much better at watching out for regime change, gauging the contextual background far better than any computer. This matters because there is ‘’strong mean-reversion in calm markets which machines must learn – and find this difficult’’. The chart of the profits based on the specific model discussed clearly show a sharp increase during this year’s market turmoil.
The key is that machine learning can provide probabilities of outcome, as well as the suggestion as to whether to bother with the trade at all. Rather tellingly he also warns: ‘’it’s no good to perform as the average trader – because most traders lose money.’’
Something I didn’t know is that deep learning’s GPT-3 has 175 billion parameters to fit into neural networks! Understandably Chan says: ‘’we are museum curators, not supermarket operators.’’
Links you might find useful: