meridianal/ARA.AI

Meridian.AI — Financial Prediction Models

Overview

Meridian.AI is a deep-learning system for predicting price movements across stocks and forex pairs. Version 5.1 keeps the MeridianModel architecture introduced in v5.0 (Grouped Query Attention + Mixture-of-Experts + optional Mamba SSM) and adds a hardened training pipeline: signal-safe shutdown, atomic checkpoint writes, comprehensive Comet ML telemetry on every run, and an audit trail of which symbols / date ranges fed each training job.

Repository layout

meridianal/ARA.AI/
├── models/
│   ├── Meridian.AI_Stocks.pt    ← current v5 stock checkpoint
│   └── Meridian.AI_Forex.pt     ← current v5 forex checkpoint
└── legacy/
    └── <archived pre-v5 checkpoints>

Anything older than v5 has been moved into legacy/. Loaders accept both old (RevolutionaryFinancialModel-2026) and new (MeridianModel-2026) architecture strings, but newly trained checkpoints always advertise MeridianModel-2026 and version=5.1.0.

Model versions

Version	Architecture string	Params	Status
v5.1	MeridianModel-2026	~11M (CPU)	Current — hardened CI, full Comet telemetry, atomic saves
v5.0	MeridianModel-2026	~11M (CPU)	Loadable
v4.1	RevolutionaryFinancialModel-2026	~45M	Loadable; archived in legacy/ on HF
≤ v4.0	various	various	Archived; loader refuses

Architecture

MeridianBlock

Each layer of MeridianModel is a MeridianBlock containing:

1. RMSNorm — pre-norm before attention
2. GroupedQueryAttention (GQA) — multi-head attention with fewer KV heads; QK-Norm for stability; RoPE position encoding
3. Optional MambaBlock — vectorised selective-scan SSM (disabled in CPU default)
4. Layer Scale — per-block learnable scalar (init 0.1) for training stability at depth
5. Stochastic Depth — drop-path regularisation
6. RMSNorm — pre-norm before MoE
7. MixtureOfExperts — num_experts SwiGLU expert networks, top-2 routing

Component table

Component	Implementation	Purpose
Attention	GQA + QK-Norm	Reduced KV cache, training stability
Position	RoPE	Relative temporal awareness
Expert routing	MoE, top-2, SwiGLU	Regime-specific specialization
Activations	SwiGLU	Better gradient flow vs GELU/ReLU
Normalisation	RMSNorm + Layer Scale	Stable training at depth
Regularisation	Stochastic Depth	Generalisation, prevents overfitting
Optional SSM	Mamba (vectorised scan)	Long-range sequential dependencies
Loss	BalancedDirectionLoss	Joint regression + direction accuracy

Model specifications

Spec	CPU Default (v5.0)	GPU / Large
Parameters	~11M	~45M
Hidden dimension	256	384
Layers	6	6
Attention heads	4	6
KV heads	2	2
Experts	4 (top-2)	4 (top-2)
Prediction heads	4	4
Mamba SSM	Disabled	Optional
Input features	44	44
Sequence length	30 timesteps	30 timesteps

Available models

Meridian.AI Stocks

• File: models/Meridian.AI_Stocks.pt
• Coverage: 49+ equities — AAPL, MSFT, GOOGL, AMZN, TSLA, NVDA, JPM, SPY, and more
• Data: Daily + 2yr hourly + 5yr weekly OHLCV with 44 technical indicators
• Training: Automatic CI retrain (GitHub Actions)
• Tracking: Comet project meridianalgo/meridian-ai-stock-v5

Meridian.AI Forex

• File: models/Meridian.AI_Forex.pt
• Coverage: 22 currency pairs — EUR/USD, GBP/USD, USD/JPY, AUD/USD, USD/CHF, USD/CAD, etc.
• Data: Multi-timeframe OHLCV with 44 technical indicators
• Training: Automatic CI retrain (GitHub Actions)
• Tracking: Comet project meridianalgo/meridian-ai-forex-v5

Usage

from huggingface_hub import hf_hub_download
from meridianalgo.unified_ml import UnifiedStockML

model_path = hf_hub_download(
    repo_id="meridianal/ARA.AI",
    filename="models/Meridian.AI_Stocks.pt"
)

ml = UnifiedStockML(model_path=model_path)
prediction = ml.predict_ultimate("AAPL", days=5)
print(prediction)

from huggingface_hub import hf_hub_download
from meridianalgo.forex_ml import ForexML

model_path = hf_hub_download(
    repo_id="meridianal/ARA.AI",
    filename="models/Meridian.AI_Forex.pt"
)

ml = ForexML(model_path=model_path)
prediction = ml.predict("EURUSD=X", days=5)
print(prediction)

Training configuration

Setting	Value
Optimizer	AdamW (weight_decay=0.02, betas=(0.9, 0.95))
LR warmup	2-epoch linear ramp, 10% → 100% of base LR
Scheduler	CosineAnnealingWarmRestarts after warmup
Loss	BalancedDirectionLoss (60% Huber + 40% weighted BCE)
Effective batch size	256 via gradient accumulation
Gradient clipping	Max norm 1.0
EMA	Decay 0.999 — used for validation and saved checkpoint
Data augmentation	Gaussian noise (0.5%) + timestep masking (5%)
Early stopping	Patience 15 on EMA validation loss
Mixed precision	bfloat16 on CPU, float16 on CUDA
Feature clamping	[-10, 10] after z-score normalisation
Sample cap	60K most-recent rows per run
CI budget	35 minutes, up to 999 epochs (was 45 — tightened in v5.1)
Signal handling	SIGTERM/SIGINT → save best EMA, then exit (v5.1)
Checkpoint write	Atomic (.tmp → os.replace) — never partial (v5.1)
Comet logging	Every loss, every metric, per-symbol dataset audit (v5.1)

Checkpoint format

Every v5.0 .pt file contains:

{
    "model_state_dict": ...,       # PyTorch weights
    "model_type": "stock",         # or "forex"
    "architecture": "MeridianModel-2026",
    "version": "5.1.0",
    "input_size": 44,
    "seq_len": 30,
    "dim": 256,
    "num_layers": 6,
    "num_heads": 4,
    "num_kv_heads": 2,
    "num_experts": 4,
    "num_prediction_heads": 4,
    "dropout": 0.1,
    "use_mamba": False,
    "mamba_state_dim": 4,
    "scaler_mean": Tensor,         # shape (44,)
    "scaler_std": Tensor,          # shape (44,)
    "metadata": {
        "best_val_loss": float,
        "direction_accuracy": float,   # percent (0–100)
        "target_min": float,
        "target_max": float,
        "training_history": [...],
    }
}

Technical indicators (44 features)

Category	Indicators
Price	Returns, Log Returns, Volatility, ATR
Trend	SMA (5/10/20/50/200), EMA (5/10/20/50/200)
Momentum	RSI, Fast RSI, Stochastic RSI, Momentum, ROC, Williams %R
Oscillators	MACD, MACD Signal, MACD Histogram, Stochastic K/D, CCI
Volatility	Bollinger Bands (Upper/Lower/Width/%B), Keltner Channels (Upper/Lower/%K)
Volume	Volume SMA, Volume Ratio, OBV (normalized)
Trend Strength	ADX, +DI, -DI, Price vs SMA50/SMA200, ATR%
Mean Reversion	Z-Score (20d), Distance from 52-week High

Limitations

1. Performance may degrade during black swan events or extreme market dislocation.
2. Predictive accuracy decreases as the forecast horizon extends beyond a few days.
3. The model reflects statistical patterns in historical data — these patterns may not persist.
4. Pre-v5.0 checkpoints (before 2026-05-15) have a known bug where validation never ran; direction accuracy is unreliable on those models.
5. For research and educational use only — not financial advice.

Citation

@software{meridianalgo_2026,
  title  = {Meridian.AI: Financial Prediction Engine},
  author = {MeridianAlgo},
  year   = {2026},
  version = {5.1.0},
  url    = {https://github.com/MeridianAlgo/AraAI}
}

Disclaimer

These models are for research and educational purposes only. They do not constitute financial advice. Trading financial instruments carries significant risk. Past performance does not guarantee future results. The developers and contributors are not liable for any financial losses. All trading decisions are yours alone.

License

MIT License. See the GitHub repository for details.