AI Model Limitations

Tradewink routes AI work through OpenRouter to models from several providers — currently including Google (Gemini), DeepSeek, Qwen, and OpenAI (GPT), among others — and also supports bring-your-own-key access for OpenAI, Anthropic (Claude), Google, and OpenRouter. The specific model assigned to a task varies by subscription tier, request type, and availability, and the roster changes over time. No single model is treated as authoritative.

Models are general-purpose large language models. They are not purpose-built financial forecasting systems. They have been trained on public data with a knowledge cutoff, which means they have no access to real-time information unless we explicitly feed it to them as context. Any market data, news, filings, or price history cited in an analysis came from our data providers (Polygon, Finnhub, SEC EDGAR, etc.) — not from the model's training data.

LLMs can and do generate plausible-sounding but factually incorrect statements. This is a known, well-documented property of the underlying technology. It is not a bug we will eliminate — it is a characteristic of the tool. We mitigate it by grounding prompts in structured data from trusted providers, by using multi-agent debate for high-stakes decisions, and by logging every AI response for post-trade review.

Despite those mitigations, an AI-generated analysis may contain incorrect ticker references, wrong dollar amounts, fabricated news events, misremembered earnings dates, or invented company names. You should never place a trade based solely on an AI explanation without verifying the underlying data yourself.

Tradewink publishes backtest and paper-trading results for several strategies and LLM benchmarks. These results are historical simulations, not promises about future performance. They are subject to the standard limitations of hypothetical results: they benefit from hindsight, they can be overfit to the specific period tested, they do not account for the full cost of real-world slippage and partial fills, and they assume execution at prices that may not have been available in size.

Survivorship bias is a real factor. Strategies that looked good in backtests have repeatedly failed in live trading. A strategy that worked for five years can stop working tomorrow if the underlying market regime changes. Past performance is not an indicator of future results — not because the phrase is required by regulators, but because it is literally true.

Tradewink includes an adaptive strategy selector — a contextual multi-armed bandit (UCB-Tuned by default, with Thompson Sampling as an alternative mode; often loosely called 'reinforcement learning', though it is not deep RL) that reweights a set of rule-based strategy families based on recent performance, conditioned on market regime and time of day. It also includes an ML retrainer that refits gradient-boosted models on a rolling window, and a dynamic exit engine that adjusts stops in real time. We describe this as 'self-improving' because the system updates itself without manual intervention.

It does not mean the system gets monotonically better over time. A model that retrains on recent data can overfit to a regime that has already ended. A bandit selector can get trapped favoring a strategy whose edge has already decayed. An auto-tuned stop can be too tight in a volatile session. We guard against the worst failure modes with validation gates, walk-forward testing, and manual review of new strategies before they go live — but we cannot guarantee that any update will improve results.

Many signals and trade evaluations include a confidence score (typically 0-100). This score reflects the AI's internal assessment of how well the setup matches its criteria — it is not a probability of profit. A signal with 95% confidence is not 95% likely to be profitable. It is a signal where the AI judged its own criteria to be strongly met.

High-confidence signals still lose. Low-confidence signals still win. The score is useful for ranking and filtering, not for predicting individual trade outcomes.

Public subscriptions are paper-only; separately approved private beta accounts may submit live broker orders. The system can identify a candidate, generate an analysis, and, with Paper Autopilot turned on, simulate the trade in a simulator or a paper/sandbox account. Paper Autopilot is off by default. For public-plan users, any real trade is your own decision, placed by you at your broker outside Tradewink, and you are responsible for it.

We designed the product this way on purpose. The AI is best treated as a tireless research assistant that scans markets you do not have time to watch, surfaces setups you would not have found, and drafts a rationale you can accept or reject. The final decision should remain yours.