AI second opinion: a second brain next to the engine
What the AI card is, what the model sees (and does not see), how the two opinions are compared, how both are scored afterwards, costs and limits.
Two brains, one decision
The engine's consensus call is statistical: it fires only when technique combinations that were profitable on this pair's own history (after fees and funding) are firing again, and it sizes entry, stop and target from what those combinations did in the past. That is a strength (every number is backed by cases) and a blind spot (it cannot reason about a situation it has not seen in numbers).
The AI second opinion adds a second, independent brain. When you press Ask AI on a pair page, the platform sends a compact market briefing to the LLM the admin selected (Claude, GPT, Gemini, Grok, DeepSeek, a local Ollama model and more) and asks for its own view: direction, confidence, horizon, entry / stop / target, three to five reasons, the key risk and what would invalidate the call. The model answers through a fixed form, so the result is always machine-readable.
What the model sees, and what it never sees
The briefing contains the last 24 candles on 15m, 1h, 4h and 1d, the indicator readings (RSI, ADX, MACD, Bollinger position, distance from the moving averages, volume and order-flow readings, trend score), the pattern bias per timeframe, derivatives data (funding and its z-score, open-interest changes, long/short and taker ratios, liquidations), order-book imbalance, the crash-risk reading and the macro calendar window.
It never sees the engine's own call. That is deliberate: a second opinion that knows the first opinion is not a second opinion. Both views are produced from the same data at the same moment and only then compared.
How the two are compared
| Engine | AI | Card badge | Combined decision |
|---|---|---|---|
| long / short | same direction | AGREE | Trade it: full size, engine's entry / stop / target; the AI's stop is a sanity check. Combined confidence = 1 − (1 − engine)(1 − AI). |
| long / short | opposite direction | CONFLICT | WAIT. Stand aside, or at most half size for the side with the better track record on this pair. |
| long / short | neutral | ENGINE ONLY | Lean (half size) in the engine's direction; watch the AI's key risk. |
| neutral / WAIT | long / short | AI ONLY | WAIT: paper-trade the idea or wait for the engine to confirm. |
| neutral | neutral | BOTH WAIT | No trade. |
The decision text also quotes the track record on this pair once at least ten opinions are scored.
Scoring: who was right
Every opinion is stored with the price at the time. Once the AI's horizon has elapsed (15 min to 1 day), the platform looks at the realised move: a directional view is right when the close moved its way by more than a third of an ATR, a neutral view is right when price stayed inside that band. The card shows the outcome (return, band, best and worst excursion, whether the stated stop or target was touched) and the scoreboard: how often the engine was right, how often the AI was right, and how both did when they agreed. A disagreement that keeps resolving in the AI's favour is a hint that the engine's history on that pair has gone stale (run a deep analysis); the opposite is a hint to weight the AI less.
Which AI? The admin chooses
The platform is not tied to one vendor. Under Settings › AI second opinion the admin picks the provider and stores that provider's key (keys are kept per provider, encrypted, so switching back and forth never loses one):
| Provider | Default model | Notes |
|---|---|---|
| Anthropic (Claude) | claude-opus-5-5 | best reasoning; Sonnet is cheaper |
| OpenAI (GPT) | gpt-5 | |
| Google (Gemini) | gemini-2.5-pro | |
| xAI (Grok) | grok-4 | |
| DeepSeek | deepseek-chat | very cheap |
| Groq | llama-3.3-70b-versatile | fast open models |
| Mistral | mistral-large-latest | |
| OpenRouter | any | one key, hundreds of models |
| Ollama | llama3.1 | runs on your own machine, free, no key |
| Custom | — | any OpenAI-compatible endpoint (base URL + model) |
The model id is free text, so a new model works the day it is released. Every stored opinion records which provider and model answered, so the scoreboard tells you whether a switch helped.
Does the AI trade? Whose call does automatic trading follow?
By default the engine alone trades. The AI is information on the card (advisory role). Two stricter roles exist under trading role of the AI:
- confirm: an automatic trade opens only when the AI, asked at that moment, independently calls the same direction. No answer (no key, provider down, cap reached) means no trade.
- veto: an automatic trade opens unless the AI calls the opposite direction. Neutral or no answer lets the engine trade.
In every role the engine's signal is what starts a trade: the AI can stop one, never start one. In Ask me mode a proposal is always created and carries the AI verdict ("AI agrees" / "AI DISAGREES" with the reason), so you decide. A blocked automatic trade produces an alert naming the reason.
Switching it off
When the AI mostly repeats the engine on your pairs (the scoreboard shows that), switch it off with the toggle on the card header (admin) or under Settings. The engine alone decides; stored opinions and the track record stay.
Costs, limits, automatic mode
This is the only paid external service in the platform (Ollama excepted). Each opinion costs a few thousand input tokens and a couple of hundred output tokens at the provider (roughly a few cents on a frontier model, a fraction of a cent on DeepSeek or Sonnet-class models). The admin stores the key once (encrypted in the engine database) and sets:
- daily cap across the platform and per user per day (hosted mode);
- a reuse window: an opinion on the pair younger than this (default 5 min) is shown to everybody instead of a new call;
- a cooldown per pair for repeated Ask again clicks;
- Ask automatically on every new engine call: one opinion whenever the consensus turns into a LONG or SHORT (counts against the cap; opinions arrive on the card live).
On the hosted platform asking is a Pro / Elite feature (ai_opinion); reading a stored opinion is free on every plan. Everything else in the platform stays free and local: no key, no AI card, nothing changes.