A financial model can predict a market move or flag a transaction, but its output is not yet a decision. That output is a signal: information that a strategy can assess against business goals, risk appetite, policies, and regulatory requirements before anyone acts. section 1.4 section 1.4.3
A signal moves through an application in stages:
- Curated data feeds a predictive or generative model. Its output becomes an input to strategy and monitoring rules, which govern what reaches an application workflow. section 1.4
- Evaluate the output for its intended use. Credit models may use a K-S score, portfolios a Sharpe ratio, and fraud detection precision and recall; thresholds must also account for policy and operational capacity. section 1.4.3
- Check the resulting decision against real outcomes. Feedback can prompt changes to data, models, thresholds, or operating rules. section 1.4
Two market signals
A random-forest model trained on exchange-traded fund price and volume features produces up, down, or neutral market-direction predictions. Across out-of-sample time-series folds, directional accuracy is 52.2% when the model predicts a direction and the market moves. That modest predictive result is not a demonstrated trading profit: transaction costs, slippage, and capital allocation still need assessment. A complementary news-analysis pipeline uses retrieval-augmented generation (RAG) to turn retrieved articles into structured JSON. In one uranium-market example, it records sentiment_score 65, supply_risk 7/10, policy_tone 6/10, key events, and reasoning. The scores are usable features; the reasoning preserves an audit trail. The numerical and news-derived signals can then be considered together in a hybrid strategy. section 10.5.5 section 11.0 section 11.4.3
Treat a signal as evidence to test, not an instruction to trade. An LLM-generated signal can contain fabricated facts or appear predictive because future information leaked into model training; RAG grounding and human oversight reduce risks without eliminating them. Outside trading, collective user research activity can also serve as a signal for investment discovery, where long-term trust matters more than clicks alone. section 11.1.1 section 13.1.1
Figure 1.3 Financial AI applications aren’t just trained models or UIs. They connect curated data assets, predictive or generative models, strategy and monitoring rules, and application workflows. Feedback from real-world outcomes flows back into the system, helping teams refine data, models, thresholds, and operational decisions over time.
The four-layer diagram locates signals between models and governed decisions, with real-world outcomes feeding back into the system. section 1.4
