In a Harvard Business School Working Paper Sean Cao, Charles C. Y. Wang and Yi Xiang from the universities of Maryland, Harvard and The Hong Kong Polytechnic uncover a strange bias in AI-generated financial predictions.
We know from prior research investors tend to favour and be more optimistic about companies in their home markets. We don’t know exactly why this is (information advantage, familiarity, bias) but the effect is real.
The researchers wanted to know how LLMs, in this case the US-crafted ChatGPT and the Chinese DeepSeek, would lean when asked to analyze the same set of companies. Since these models are products of ‘home-teams’ in each case then a home-bias was to be expected.
That’s not what came out though. ChatGPT was consistently more optimistic about prospects for Chinese companies than DeepSeek. Both models reached the same conclusions about US companies however. So what’s going on?
Three factors are of particular note:
- LLMs seem to have a default bias to positivity when data is scarce. Because ChatGPT missed negative information available to DeepSeek it filled the holes with sunshine.
- The difference between the models’ predictions are not a function of the respective models. The difference is entirely attributable to data asymmetry. This systematic error can therefore be rectified.
- For rectification to be effective however the bias has to be recognized, which may not happen quickly, or at all. Thus, capital allocation using LLMs may increase cross border asset pricing inefficiency.
The paper concludes with a homily about how researchers should be aware that rather than reflecting or amplifying human biases AI can create new ones.
AI in this regard may not act as a democratizing valuation tool it may in fact end up exacerbating inefficiencies and “..understanding and correcting these biases becomes essential for maintaining market efficiency and protecting investors.” Not sure if a ‘D’uh!’ or an ‘Amen!’ is required here?
Either way, we should consider ourselves warned. You can access the paper in full via this link When LLMs Go Abroad.
Happy Sunday