AI bias is basically induced by either flaw in data or bias humans have as programmers.
Nothing can give jittery experience than a stock market meltdown to an investor, trading with the best possible bets. History is ripe with such examples and there is very much probability that they will repeat, for different reasons though. Trading in stocks is more than a guessing game, now partially taken over by bots, not even sparing the well-grounded players at Wall-street. It was robotic trading that caused the famous stock market crash of 1987 when Dow Jones plunged to 22.6%. 2010 the sterling ‘flash crash’ is believed to be the generous and intentional work of a rogue bot. Like with any technology, AI has its teething problems with fintech. One such often ignored but the important issue is AI bias, which can induce unnecessary and unfair outcomes. Now that, stock trading bots have become a quintessential part of the trading business, it is essential to understand what exactly induces AI bias and the ways to address them.
AI is not biased; humans who are at the helm of designing AI algorithms are:
Algorithms depend on big data and the pattern generated thereby or the way they are programmed. As Cathy O’Neil, mathematician and author of Weapons of Math Destruction said: “I know how models are built because I build them myself, so I know that I’m embedding my values into every single algorithm and I am projecting my agenda onto those algorithms.”
Traders are humans, sometimes driven by gut instincts, the resulting action fitting nowhere into a pattern. For example, a trader may decide to hold back certain shares to sell them at a time he thinks is appropriate and yields profits. This particular instance is registered as an event with the potential for a successful investment strategy in the database. When an algorithm is back-tested for identifying such events, it would definitely hold back some shares, which may not result in giving similar results.
How do unanticipated contexts and flaws in data induce biases?
Trading in stocks or rather any financial decision basically depends on crunching numbers to arrive at the right decision. The result can shift dramatically for a slight variation or misrepresentation of data. Understanding data representation, and the bias range induced thereby, is important for intelligent bots to behave. The biases can occur at any stage of algorithmic development.
Data mining bias, for example, occurs when the sample data mined is from a random incident. An algorithm can consider it as an event holding an opportunity and keep digging for such data repeatedly. In stock markets, such random and hunch-based transactions happen quite frequently. Experts, therefore, are of opinion that quality data, both in structured and unstructured formats is necessary to prevent data-mining bias.
In certain cases, when the algorithm depends on data that is not present at the time of trading such as betting on the closing price for intraday trading, also called the look-ahead bias creeps in. Data optimization through parameter adjustment can make a data set attractive; but when implemented, the strategy may not perform the way it was intended to. In the case of long-term data analysis, algorithms may not be able to identify data gaps, for example, gaps created by dysfunctional stocks. This is called survivorship bias, which happens because most trading indexes stop reporting the fallen funds.
In view of Artificial Intelligence becoming a necessary and inevitable part of the financial sector, it is imperative that we humans as AI designers should take all measures to not let a bias, either inherent or human-generated, determine the fate of fortunes.