Beyond Traditional Financial Ratios: Sentiment Analysis Becomes a New AI Investing Tool
In quantitative investing, traditional multi-factor models have long relied on rigid financial metrics like P/E, P/B, and ROE. However, with the explosive maturity of Natural Language Processing (NLP) technology, Singapore's AI investment community is undergoing a paradigm shift from 'reading numbers' to 'reading tone.' On August 4, 2026, Temasek Polytechnic's AI and Data Analytics Centre released a significant research outcome: an earnings sentiment analysis model named 'Sentiment-Quant 2026.' In historical backtests on Singapore Exchange (SGX) Mainboard-listed companies, the model achieved a striking 78% accuracy in predicting the direction of stock price movements over the next quarter, far exceeding the baseline of traditional financial ratio models.
Decoding Management's Subtext: How NLP Captures 'Alpha in Words'
Traditional earnings analysis often stops at extracting structured data like revenue, profit, and liabilities. However, Temasek Polytechnic's team discovered that a wealth of unstructured sentiment information is hidden within the 'Management Discussion and Analysis' sections of financial reports and earnings call transcripts. This includes management's confidence in the future, subtle wording on risks, and deliberately avoided questions.
The 'Sentiment-Quant' model utilizes a large language model based on the Transformer architecture, fine-tuned specifically on a professional financial corpus. The model not only identifies positive and negative words but also uses contextual semantic understanding to accurately judge management's 'vague tone' and 'false optimism.' For instance, when a company repeatedly emphasizes that 'the macro environment is challenging, but we are confident,' the AI, based on historical data comparisons, recognizes that such phrasing often precedes lower-than-expected capital expenditure cuts, thus assigning a rating below market consensus.
Live Backtesting: Outperforming the Straits Times Index in the Singapore Market
To validate the model's effectiveness, the research team used financial reports from the top 100 SGX Mainboard companies by market capitalization between 2018 and 2025 as training and testing samples. Results showed that an AI stock-picking strategy incorporating the sentiment analysis factor achieved an annualized return of 12.3% over seven years, significantly outperforming the Straits Times Index (STI) over the same period. More crucially, the model excelled at identifying 'crash-prone stocks.' During 2024-2025, it successfully warned of several small and medium-sized local Singaporean listed companies that later saw their stock prices plummet due to supply chain disruptions or debt defaults, with the average lead time for warning signals reaching 45 days.
From Institutions to Retail: Robo-Advisors Poised for a 'Sentiment-Aware' Upgrade
Following the announcement of this technology, Singapore's FinTech community reacted swiftly. Several robo-advisory platforms licensed by the Monetary Authority of Singapore are reportedly in talks with Temasek Polytechnic to integrate the 'Sentiment-Quant' model into next-generation digital wealth management tools. This means that in the future, ordinary retail investors might simply tap their mobile app to see an automatically generated 'Earnings Sentiment Score.'
The CTO of a prominent Singapore-based robo-advisor stated: 'In the past, it was difficult for retail investors to analyze earnings call transcripts word-for-word like professional institutions. Now, AI democratizes this analytical capability. Robo-advisors incorporating sentiment analysis will no longer just allocate ETFs based on risk appetite; they can provide deeper, stock-level risk avoidance advice, fundamentally changing the depth of AI-driven financial services.' This trend aligns with the AI financial regulatory sandbox policies actively promoted by the Monetary Authority of Singapore in recent years, signaling a new explosive phase for AI applications in wealth management and risk warning.
Challenges and Outlook: The 'Black Box' Risk of Sentiment Analysis
Despite the promising outlook, AI sentiment analysis still faces many challenges in practical application. Some portfolio managers worry that over-reliance on AI's interpretation of textual sentiment could lead to a 'herding' risk in trading. Furthermore, with management language styles varying widely, whether AI can truly understand subtle sarcasm or blindness born of extreme confidence remains a key debate in academia.
Temasek Polytechnic's R&D team stated that the next phase will involve introducing multimodal data, such as analyzing executives' vocal tone during earnings calls, to further improve prediction accuracy. Concurrently, the team is developing an explainable AI interface designed to clearly show users the reasons behind a model's bullish or bearish rating on a particular report, thus preventing the AI from becoming an uninterpretable 'black box.' As AI-driven financial advisory becomes increasingly prevalent, this Singaporean research undoubtedly sets a new benchmark for AI-driven fundamental analysis.
