Before this version, the AI Analysis button on a news headline or scheduled event analysed that single datapoint in isolation. The 2026-05-14 reproducer made the limit obvious: a weak US Retail-Sales print read as USD-bearish, but the previous-day hot PPI (+71% surprise), an active trade-deal narrative, and a four-day downtrend in GBP/USD all pulled the other way — the single bullish-on-paper read drove a stop-out.
News-AI v2 assembles a context bundle around every event before the prompt goes to Anthropic. The bundle contains:
- Prior events (last 7 days, high/medium impact) — name, country, actual vs. forecast, surprise percentage, and a `hot / cold / in_line / neutral` direction label that translates the surprise into a home-currency tilt.
- Recent headlines (last 48 hours) — top 5 cached news items, with cached sentiment labels where available.
- Symbol bias — 5-day return, 14-day RSI, 20-day annualised volatility, and a coarse `bullish_strong / bullish_mild / neutral / bearish_mild / bearish_strong` trend label.
- Macro proxies (USD pairs only) — 5-day change in the US 10Y yield, 5-day change in the dollar index DXY, and the current VIX level.
- Active narrative tags — keyword-matched themes derived from the headlines (`trade_deal_optimism`, `hawkish_fed`, `recession_fear`, etc).
The Anthropic prompt now asks the model to weigh the event signal against these competing drivers and respond with four structured fields:
- `isolated_signal` — one sentence on what the event says on its own.
- `competing_drivers` — up to five concrete forces that could overwhelm the isolated read.
- `net_confidence` — a 0–1 score for how much the isolated signal will dominate.
- `decision_text` — two to three sentences on what a trader should keep in mind.
The new fields render directly on the news, events, briefing, and market-heat surfaces under the existing analysis pane.
- Red banner (net_confidence < 0.5) — the event signal is in conflict with the surrounding context. Two or three other forces are pulling against it. Treat the headline read as the minority position; the rest of the tape is more likely to dominate.
- Amber banner (0.5 ≤ net_confidence < 0.7) — context is mixed. Some drivers align with the event, some push back. Watch for the deciding catalyst rather than acting on the event alone.
- Green banner (net_confidence ≥ 0.7) — the event signal stands up to context. Prior events, narrative tags, and the symbol's bias all point in the same direction. The headline is reinforcing what's already in motion.
- Competing drivers section is concrete on purpose — `Hot PPI surprise yesterday (+71%)` beats `inflation pressures`. The model is asked to name the driver and quantify where it can.
- `Isolated signal` is the legacy read — keep it for context; the `Trader read` below the drivers is the integrated take.
- A 0.3 confidence is not a sell signal — TradeOnyx never tells you to buy or sell. The score tells you the event-vs-context tension; the trading decision is yours.
Open a freshly-broken headline on the News tab and click AI Analysis. The classic two-sentence read still appears in the gold-bordered pane. Below it, the new banner tells you in two lines what the model thinks about the news vs. the surrounding macro environment.
If the banner is red, expand Competing drivers before you act. The list will name the prior event, the active narrative, or the symbol-bias that's pulling against the news read. Often the right reaction is to NOT trade that headline — the cumulative push from the other side is bigger than the new datapoint.
If the banner is green, the new headline is a confirmation, not a fresh signal. Most of the move is already priced in; this is the time to manage existing positions, not to chase a fresh entry.
The same flow applies to scheduled events on the Economy tab and to the daily Briefing read. A high-confidence briefing means today's calendar is reading with the tape; a low-confidence briefing means the calendar and the tape disagree, and the trader needs to pick which signal to follow.