BEN RisQuant schreef op 6 februari 2019 17:21:
If investors care only about profits, why should anyone worry about theory?
First, arbitrage forces subdue the signal-to-noise ratio in financial data sets. As a consequence, these data sets are noisier than natural data sets. In the absence of theories, ML algorithms confound noise with signal, a phenomenon known as overfitting. An overfit algorithm makes wrong recommendations that lead to uncompensated risks.
Second, financial systems are dynamic. Patterns may cease to repeat. A theory tells us why X causes Y, and this is helpful for determining whether the strategy is real and persistent. A financial theory gives us the confidence to invest in or stick with a strategy, even as it undergoes a drawdown.
By embracing the black box approach, investment companies may be trying to monetise the hype generated by recent ML successes. It is important for investors to recognise the fundamental differences between the two ML paradigms, and why financial black boxes are not aligned with their investment goals.
In the words of economist Lars Hansen, a Nobel laureate: “[D]ata seldom, if ever, speaks for itself.” Developing investment strategies purely by data-mining will lead to unreliable outcomes and losses, because correlations are often spurious and correlation does not imply causation. We need ML to develop better financial theories and we need financial theories to restrict ML’s propensity to overfit. Without this theory-ML interplay, investors are placing their trust on high-tech horoscopes.