Turning market signals into research-grade decision support

Client Industry

Finance · investment research team

Key Challenge

Fusing noisy technical indicators & unstructured sentiment

Core Tech Stack

Time-series · Deep learning · Sentiment analysis

Deployment Model

Production decision-support system · human-owned

SoftmaxAI has developed a research and decision-support system of production quality which assists an investment research team in reasoning about equities in a more systematic manner. The system combines two very different kinds of signal — quantitative technical indicators obtained from market price data and qualitative market sentiment extracted from news headlines and social posts — into potential directional signals that the team can consider together with their own analysis. The system is intentionally presented as a means to aid human judgment, not as a source of financial advice or as one that guarantees any outcomes; the analysts retain responsibility for all their decisions.

• Signal was buried in noise. Technical indicators are notoriously noisy, and the movements of equity prices are influenced by a complex mix of rational fundamentals and irrational crowd psychology — no single indicator can by itself tell a reliable story.
• Sentiment was unstructured. The attitude of the market was reflected in the news headlines and social media posts in written form, a type of data which quantitative models cannot take in unless they have dedicated language understanding.
• Two data worlds had to meet. It is a real modeling problem to combine structured time-series indicators with unstructured text sentiment into a single coherent view, rather than just joining them.
• Trust had to be earned, not assumed. If a signal is to be of any use, an analyst must know how it was produced and where it fails — a black box which simply states a direction is no better than having no tool whatsoever.

The system operates as a left-to-right research pipeline in which two input streams supply data to the feature engineering stage, the parallel forecasting and sentiment models are combined to produce candidate signals, and an evaluation loop on held-out data keeps the modeling component honest before the final decision is made by a human.

• Spec before code. In accordance with the SoftmaxAI Method, we settled on a written specification in advance, outlining the models, the data sources, and the evaluation criteria, before any actual production code was written.
• Parallel models per signal type. Senior engineers carried out a build that was accelerated by agents, applying time-series and deep learning methods to the technical indicators and a sentiment-analysis model to news and social text, ensuring that each of the streams was dealt with in a way appropriate to it.
• Fusion into candidate signals. The output from the model is combined into a set of candidate directional signals which appear as decision support — being presented to the analyst as possibilities to consider, not as instructions to take action.
• Measured against held-out data. Each modeling decision was examined by people and the system was assessed using data that had been set aside beforehand, which allowed the team to understand both how the system behaved and what its limitations were before it was ever considered for making a decision.

• A systematic way to weigh signals. The team acquired a disciplined and data-based way of looking at equities — an organized alternative to relying solely on intuition when assessing technicals and market mood.
• Technical and sentiment signals, unified. The quantitative indicators that are noisy and the sentiment derived from unstructured text are combined together to provide a single, coherent view which an analyst can read and reason about.
• Candidate signals that feed human analysis. The system produces candidate directional signals which are incorporated into the team's own research and strategies rather than replacing them.
• Understood behavior and limits. The team knows how the system performs and where it is unreliable since the models were evaluated using held-out data — this knowledge forms the basis for using the system responsibly.
• A decision aid, firmly human-owned. By presenting itself as a source of support rather than offering advice, the system ensures that the analysts remain in control of each call — there are no guaranteed results, no automated trades, and no financial advice.