A new artificial-intelligence investment simulator could make stock-market algorithms less mysterious by showing users not only what an AI recommends, but also which signals influenced its decision and how long a desired return might take to emerge. The system, developed by researchers at Complutense University of Madrid, combines deep-learning agents with explainable artificial intelligence in an interactive environment called the Dashboard for Intelligent Investment with XAI, or DI2XAI. In a study involving 51 participants, the platform helped users improve their profitability relative to the market between their first and final simulation sessions. The researchers report an average relative improvement of 8.71 percent, suggesting that transparent AI tools may change how people learn to interpret automated trading decisions rather than simply encouraging them to follow algorithmic instructions.
Stock investment is an especially difficult arena for machine learning because financial data are noisy, rapidly changing and shaped by events that are not always represented in historical price records. A model can identify statistical patterns in prices, trading indicators or other market variables, but those patterns may disappear when economic conditions change. The problem is compounded by the fact that many high-performing models are difficult to interpret. Deep neural networks can contain large numbers of adjustable parameters, allowing them to capture complicated relationships while making it hard for a human user to understand why a particular buy or sell signal appeared. DI2XAI was designed to address both challenges by pairing predictive agents with visual explanations and by allowing users to test decisions against a long historical period rather than a narrow slice of market history.
The platform uses two types of machine-learning model: Long Short-Term Memory networks, or LSTMs, and Multi-Layer Perceptrons, or MLPs. An LSTM is a recurrent neural-network architecture designed to process sequences. Its internal memory mechanisms allow it to retain or discard information across successive time steps, making it suitable for data in which recent price movements may depend on earlier trends. An MLP, by contrast, is a feed-forward network composed of connected layers of artificial neurons. It receives input features and transforms them through weighted mathematical operations and nonlinear activation functions before producing an output, such as a predicted market direction or trading action. Within DI2XAI, these models operate as investment agents that make decisions in a simulated brokerage environment, giving users the opportunity to observe and challenge their behavior.
The crucial interpretive layer comes from SHAP, short for Shapley Additive Explanations. The method is based on a concept from cooperative game theory known as a Shapley value, which estimates how much each feature contributes to a model’s output by examining its marginal contribution across combinations of inputs. In a trading interface, a SHAP explanation might indicate that a particular technical indicator, recent price movement or other input pushed an agent toward buying, while another feature weakened that recommendation. Positive and negative contributions can be displayed graphically, transforming an otherwise opaque prediction into a ranked account of the factors associated with it. The technique does not reveal a simple causal law governing the market, and it cannot guarantee that the model is correct, but it can expose the evidence the model used internally and help users detect implausible or unstable reasoning.
DI2XAI adds another feature aimed at correcting a common misunderstanding of investment predictions: the difference between reaching a return target and reaching it within a specific amount of time. The system estimates a probability distribution for the time required to achieve a chosen return under particular market conditions. Rather than presenting a single deterministic forecast, it attempts to show how likely different durations may be. This is important because a target gain of, for example, 10 percent has a very different practical meaning if it is expected within weeks, several months or an uncertain period extending years into the future. Expressing the result as a distribution acknowledges that financial outcomes are inherently probabilistic and that a model’s estimate of timing carries uncertainty of its own.
To reduce the danger of conclusions based on unusually favorable short-term conditions, the researchers evaluated the simulator through one decade of historical backtesting. Backtesting applies a trading strategy to past data to estimate how it would have performed, although it cannot reproduce every condition of live investing. A long window can expose an agent to a wider range of market environments, including sustained rises, declines and periods of instability, and may therefore provide a more demanding test than a brief sample. Yet historical simulation remains vulnerable to issues such as changing market structure, transaction costs, data-selection effects and the risk that a strategy has indirectly benefited from information that would not have been available at the time of a decision. The decade-long design is consequently best understood as an effort to broaden evaluation, not as proof that the agents can reliably predict future markets.
The researchers also examined how people responded to the platform. Participants interacted with the simulated agents and assessed the usefulness of the explanations and the duration–probability tool on seven-point scales. SHAP-based explanations received mean scores of 5.91 among participants with economic experience and 5.46 among those without it. The time-to-target analysis was rated even more highly, with mean scores of 6.45 and 5.71, respectively. Those results suggest that users found information about the likely time horizon especially valuable, perhaps because it connects an abstract model output to a practical investment question. The scores also indicate that explainability is not automatically understood in the same way by every audience: financial knowledge, familiarity with technical indicators and confidence in interpreting probabilistic information may all shape whether an explanation is genuinely useful.
Across the initial and final simulation sessions, participants increased their relative profitability compared with the market by an average of 8.71 percent. The reported Cohen’s d effect size was 0.38, generally interpreted as a small-to-moderate difference, while the study reported statistical power of 86 percent. These figures are encouraging but should not be confused with evidence that users earned real-world profits or that the AI agents possess a durable market advantage. The participants were learning within a controlled simulator, where the available information, feedback and trading environment were defined by the researchers. Improved performance may therefore reflect greater familiarity with the interface, practice in reading charts or better understanding of the simulated rules. A larger study with a control group, independent market periods and real transaction costs would be needed to determine how much of the improvement came specifically from SHAP explanations or the duration estimator.
The project’s broader significance lies in treating explainable AI as part of the decision-making environment rather than as an after-the-fact technical decoration. A recommendation that merely says “buy” or “sell” can encourage automation without understanding, while an explanation can prompt users to question which indicators matter, whether the evidence is consistent and how uncertain the prediction may be. The authors have made datasets associated with the study available through a Mendeley Data repository, and the system is designed to be extended with new indicators, trading strategies and explanation methods. That flexibility could allow researchers to compare alternative models and test whether explanations remain stable when the input data or market regime changes. At the same time, no visualization can eliminate financial risk: SHAP describes a model’s attribution of importance, not the true causes of market movements. DI2XAI therefore points toward a more transparent form of algorithm-assisted investing—one in which users can inspect, test and challenge machine-learning decisions—while leaving the central uncertainty of markets firmly in place.
