The key contributions of this work are as follows:

A modular multi-agent architecture in which four domain-specific agents share a common BaseAgent interface, enabling independent development, testing, and substitution of each component without disrupting the orchestrator.

The key contributions of this work are as follows:

The central problem with standard recurrent neural networks is the vanishing gradient: as backpropagation unfolds through time, error signals shrink exponentially, making it practically impossible for the network to learn dependencies spanning more than a handful of timesteps. Hochreiter and Schmidhuber [1] addressed this limitation with the LSTM architecture. The key innovation is the cell state—a dedicated information highway that gradients can traverse without repeatedly passing through sigmoid compressions that attenuate their magnitude.

Three gating mechanisms regulate this highway. The forget gate decides what fraction of the previous cell state to retain. The input gate controls how much of the current input is written in. The output gate determines how much of the cell state is exposed as the hidden state passed to the next timestep, as illustrated in Figure 1.

Fischer and Krauss [5] established LSTMs as the benchmark for sequential financial prediction, demonstrating consistent out-of-sample gains on S&P 500 constituent data. The implementation here stacks two LSTM layers (hidden_size=128, dropout=0.2), followed by Linear(128→64), ReLU, Dropout(0.2), and Linear(64→1). Training ran for 600 epochs with the Adam optimizer (lr=0.001) and mean-squared error loss.

Vaswani et al. [11] replaced recurrence entirely with self-attention: every position in the input sequence directly attends to every other position simultaneously, with no sequential bottleneck. The core operation is Attention(Q,K,V) = Softmax(QKᵀ / √dₖ) × V, where Q, K, and V are learned linear projections. Multi-head attention runs several such operations in parallel across different representational subspaces, then concatenates and projects their outputs

The implementation here is intentionally lightweight: a Linear embedding (1→64), two TransformerEncoderLayers (nhead=4, dim_feedforward=128, dropout=0.2), and a Linear output projection. Training used Adam (lr=0.0005) for only 15 epochs—a short run reflecting runtime constraints during development, which explains its low ensemble weight of 3.7%. Wen et al. [12] provide a comprehensive survey of transformer applications in time-series forecasting. Extended training would almost certainly increase this contribution substantially.

Chen and Guestrin [3] introduced XGBoost as a scalable, regularized gradient-boosted tree framework. It builds an ensemble of decision trees sequentially, with each tree correcting the residuals of its predecessors using a second-order Taylor approximation of the loss function. Its practical advantages in financial modelling include interpretability through feature importance scores, robustness to irrelevant features, and strong generalization

from limited labeled data. The implementation uses XGBRegressor with n_estimators=200, max_depth=6, learning_rate=0.05, and subsample=0.8.

Devlin et al. [7] demonstrated that bidirectional pre-training on large text corpora produces representations that transfer powerfully to downstream NLP tasks. Araci [8] extended this approach to the financial domain by fine-tuning BERT on the Financial PhraseBank corpus [13], producing FinBERT—a model capable of understanding the hedging constructions, conditional tenses, and domain-specific polarity markers that general-purpose sentiment models systematically misclassify. The sentiment score used throughout this paper is P(positive) − P(negative), derived from FinBERT's three-class softmax output.

Mnih et al. [9] demonstrated that deep Q-networks can learn effective policies in complex sequential decision environments using only reward-based feedback, establishing RL as a principled approach to algorithmic trading. The component used here is simpler—tabular Q-learning suits the small 8-state space created by the three binary agent signals—but follows the same fundamental principle: Q(s,a) ← Q(s,a) + α[r + γ·max Q(s',a') − Q(s,a)], with α=0.1 and γ=0.9. The theoretical foundation is provided by Sutton and Barto [10].

AAPL daily closing prices were downloaded from Yahoo Finance covering January 1, 2015 to January 1, 2024-nine full years of trading data including the COVID-19 disruption period (2020), a strong bull market (2021), the technology sector correction (2022), and the subsequent recovery and AI-driven rally (2023). AAPL was selected for its high liquidity, extensive analyst coverage, and broadly representative price dynamics within large-cap US equities.

Prices were normalized to 0, 1 using MinMaxScaler, and sliding 60-day windows were extracted: the input to each model is the sequence [closeᵢ₋₆₀, ..., closeᵢ₋₁], and the target is closeᵢ. A strict 80/20 chronological split—with no shuffling-produced 1,763 training samples and 441 test samples covering January 2020 to January 2024. Shuffling was deliberately avoided to prevent look-ahead leakage.

Table 1: Model Architectures and Training Configurations.

Ensemble weights are derived from observed test-set RMSE values. For model i, the weight is its normalized accuracy: wᵢ = (1/RMSEᵢ) / Σⱼ(1/RMSEⱼ). This formulation gives higher-accuracy models proportionally greater influence while still allowing all models to contribute. The resulting weights from evaluation are LSTM = 0.6252, XGBoost = 0.3377, and Transformer = 0.0371. The Transformer's near-zero weight reflects its abbreviated training, not an inherent architectural limitation.

All four agents inherit from a BaseAgent abstract class that enforces a single required method: act(state). This minimal interface makes the architecture modular by design—any agent can be replaced, upgraded, or tuned in isolation without touching the orchestrator or the other agents. Fig. 3 illustrates the complete multi-agent system, and Table II details each agent's inputs, internal logic, and outputs.

Table 2: Agent inputs, internal logic, and outputs.

After the Decision Agent produces its majority-vote recommendation, the reinforcement learning layer observes the three-bit state—(market_signal, sentiment_signal, prediction_signal)—and selects the action with the highest Q-value for that state. The reward signal is the realized price change in the direction of the chosen action: positive when the system is correct, negative when it is wrong. The agent trains through the full 453-step test sequence, updating Q-values at each step using the Bellman equation [10].

Table III summarizes model performance on the 453-day test set. All RMSE and MAE values are reported in USD after inverse-scaling. The LSTM is the strongest individual model, achieving RMSE = 5.44 USD and R² = 0.915—meaning it accounts for over 91% of the variance in AAPL's closing price during the test period. XGBoost trails substantially at RMSE = 10.07 and R² = 0.709, reflecting its structural disadvantage on sequential prediction tasks.

The dynamic ensemble improves R² to approximately 0.96 and achieves the lowest MAE of the evaluated configurations (~$3.40), confirming that model combination adds value beyond any individual predictor [5]. The higher ensemble RMSE relative to the standalone LSTM is expected: the 3.7% Transformer weight introduces variance at timesteps where the undertrained Transformer is least accurate.

Table 3: Test-set performance on aapl (jan 2020–jan 2024, 453 days) Source column shows key references supporting the model architecture used.

Table 4: Backtesting results -$10,000 initial portfolio

Table 4 reports backtesting results from a $10,000 initial portfolio evaluated over the four-year test period (January 2020 to January 2024). The system grew the portfolio to $11,566.98, a total return of 15.67%. The annualized CAGR of 8.69% and positive Sharpe Ratio of 0.6536 confirm that excess returns were generated per unit of risk taken. The maximum drawdown of −10.04% reflects the system's response to the 2022 technology sector correction.

Table 5: Finbert sentiment scores (score = p(positive)-p(negative))

Table 5 presents FinBERT sentiment scores for three representative headlines. Two of the three scores appear counter-intuitive: a headline describing rising stock prices generates a bearish score, while one describing market uncertainty generates a bullish score. These patterns are not errors—they reflect domain-specific linguistic priors acquired during fine-tuning on the Financial PhraseBank corpus [13], where phrases like 'rises after' often signal that positive news is already priced in, and uncertainty language is frequently associated with recovery forecasts.

The LSTM's directional accuracy of 49.77%-marginally below 50%-can appear alarming at first reading, but warrants careful interpretation. This metric measures whether the model correctly predicts the sign of the next-day closing price change. Daily equity direction is widely accepted in the financial literature to approximate a random walk; sustained directional accuracy above 55% net of noise is rare even among institutional fund managers [5].

The LSTM was optimized for price-level accuracy (MSE loss), not directional accuracy, and its R² of 0.915 confirms strong performance on its actual objective.

The residual plot in Figure 12 shows prediction errors (Actual − Predicted) plotted against the predicted price level. Errors are approximately zero-centred across the mid-range of $100–$160, indicating good calibration in that regime. However, above $160, a positive skew emerges: when the model predicts high prices, actual prices tend to be higher still. This is the momentum-lag problem—a well-documented limitation of recurrent models trained on smoothed sequences [5].

Figure 13: Prediction error distributions (KDE). LSTM (narrowest distribution) MAE = 4.64; XGBoost (heaviest tails) MAE = 6.87; Transformer (intermediate). Error widths directly mirror the individual MAE figures.

Figure 13 shows kernel density estimates of prediction errors for all three individual models. The LSTM distribution is the narrowest, centred closest to zero, consistent with its lowest MAE of $4.64. XGBoost shows the heaviest tails, reflecting its occasional large errors during trend reversals, consistent with its MAE of $6.87. The Transformer occupies an intermediate position despite only 15 training epochs, hinting at its architectural potential when properly trained [11,12].

Figure 14: Buy and sell signals overlaid on AAPL actual price. The system captures the 2022 correction and 2023 recovery at a macro level. False signals concentrate during sideways consolidation, a known limitation of trend-following systems.

Figure 14 overlays the system's buy and sell signals on the actual AAPL price history. The system broadly identifies the 2022 correction and the 2023 recovery correctly at the macro level. False signals cluster during sideways consolidation periods-a characteristic weakness of trend-following ensembles, which generate spurious signals when prices lack a clear directional trend [6].

No transaction costs or bid-ask spread modeling-net returns at realistic turnover would be lower. (2) Single-asset evaluation on AAPL, a highly liquid large-cap stock.

Static sentiment: one headline per inference rather than a real-time aggregated news stream.

RL sentiment signals are randomized during training, fundamentally limiting policy quality for sentiment-dependent states.

Transformer trained for only 15 epochs its ensemble weight substantially underrepresents its architectural potential [17-20].