An explainable market forecaster improves on its baseline—but gets return direction right only 52% of the time

A financial forecasting system can offer an explanation and still struggle to predict tomorrow. That tension runs through a preprint by Brandon Rodriguez, Abel Deulofeu and Olaoluwa Adigun at Florida International University. Their hybrid model correctly identified the sign of the next day’s NASDAQ return in 52.0% of test cases, compared with 44.4% for their standalone neural-network baseline. The improvement is real within the reported comparison, but it does not establish a profitable trading strategy.

The researchers investigate two ways to represent financial relationships using fuzzy cognitive maps. These are networks in which nodes represent concepts, such as unemployment, market volatility or price momentum. Signed connections describe whether one concept is modeled as pushing another upward or downward; their weights describe the strength of that influence. Unlike a fixed checklist of indicators, the map can repeatedly update its internal states through these connections.

A map that adjusts as observations arrive

The first framework combines adaptive maps with a long short-term memory network, or LSTM: a neural model designed to learn patterns in sequences. Its inputs include prices, economic indicators, technical signals and sentiment extracted from financial headlines. The pipeline draws on Yahoo and Google financial sources alongside economic data from FRED.

Headline sentiment is scored using a financial vocabulary, smoothed over time and delayed by one trading day to reduce look-ahead bias. Numerical features use a rolling normalization window of roughly one trading year. The LSTM processes sequences covering 30 trading days.

The maps also change after observations arrive. An error-driven update adjusts connections into the NASDAQ target node, with momentum carrying part of earlier adjustments forward. Connection weights are clipped to a bounded range. The authors report that the maps stabilize after market-regime changes within 10–15 trading days, although this reported adaptation speed is not itself evidence of dependable forecasting across future disruptions.

Rather than simply averaging every model’s output, the ensemble lets the maps nudge the LSTM prediction. Bullish and bearish signals receive slightly different adjustment strengths, selected using validation data. In practical terms, the neural network supplies the starting prediction and the maps provide a limited directional correction.

What the held-out test shows

The first framework uses data from 2020–2025, assigning 2020–2023 to training, 2024 to validation and 250 trading days in 2025 to testing. Its return-sign accuracy rises from the LSTM’s 44.4% to the ensemble’s 52.0%, a gain of 7.6 percentage points. The standalone maps score 48.8%, leaving the ensemble 3.2 points ahead of them.

The paper also reports a separate measure called state-directional accuracy, intended to describe bullish versus bearish market states. Table II gives 78.8% for the ensemble and 78.4% for the LSTM. These nearly equal results should not be confused with correctly predicting the sign of daily returns: the stronger-looking state figures answer a different question.

That distinction needs particular care because the paper describes state classification using a threshold of 0.5 in normalized space, while return-sign accuracy uses the zero-return boundary. The midpoint of a rolling min–max scale does not necessarily represent zero return. Readers therefore should not interpret the roughly 79% state score as a next-day up-or-down success rate.

A second map exposes the timing problem

The second framework builds a fixed map from literature-derived relationships, supplemented by GPT-4 suggestions and citation-based checking intended to support human review. Unlike the first framework, its construction does not depend on live price or sentiment feeds. Evaluation subsequently associates its nodes with market data.

Across 119 trading days from February to July 2026, four map variants achieve same-day state-directional accuracies between 68.91% and 85.71%. But with inputs delayed by one trading day, three score 50.42% and the fourth reaches 54.62%. Seven-day delays produce results around 46–50%. The high same-day figures therefore do not demonstrate equally strong advance forecasts.

The broader limitations matter as much as the improvements. The paper does not report trading returns after costs or statistical uncertainty around its headline accuracy gains. Its principal hybrid comparison covers one index and one test year. Moreover, connections initialized from co-movement or supported by literature citations remain modeling assumptions, not experimentally established causal effects.

The study offers an inspectable way to steer a neural forecast and update some relationships over time. Its results support that narrower contribution—not a claim that transparent financial maps have solved market prediction.