Methodology

Model names live here — not in the homepage pitch. The live board is horizon-specific scores, expected returns, and price targets. Research architectures are evaluated before they can serve.

What the live product publishes

TradeWell combines more than 100 proprietary composite signals spanning price, volatility, fundamentals, technicals, momentum, money flow, sentiment, news, insider activity, seasonality, macroeconomic conditions, and valuation.

Each of 2,800+ stocks receives a separate TW Quant Score, expected return, and model-implied price target for 1D, 3D, 1W, 2W, 1M, 3M, 6M, 1Y, 3Y, and 5Y. Signals and forecasts refresh daily on current market data and historical daily bars where available. Live rankers are evaluated on NDCG@10. Realized returns feed back into training.

Today’s served ensemble is the production ranker family (gradient-boosted trees, linear baselines, and Elite deep rankers). A name is not “live” because it appears in a research paper or a challenger job.

What the research program evaluates

TradeWell’s research program evaluates horizon-specific forecasts across time-series foundation models, transformers, sparse mixture-of-experts, state-space models, and graph neural networks using walk-forward out-of-sample testing — before any model can reach live rankings.

Architectures under evaluation include:

• Time-MoE
• Lag-Llama
• Temporal Fusion Transformer
• Chronos
• Mamba-based state-space models
• Dual-Graph Equity Attention Networks
• TabPFN
• Gradient-boosted ranking models

None of those names is a claim that the architecture currently serves the public board. Promotion is operator-only. Never auto-promote.