Nexus AI

How it works

Methodology, calibration, and limits

A briefing product is only as honest as its methodology page. We publish ours so you can decide how much weight to give the briefing.

The pipeline

Every Sunday at 8 PM ET, an autonomous AI agent (Anthropic Claude) is given five tools and asked to produce a structured weekly briefing on QQQ. Tools are simple, auditable, and deterministic where possible:

get_prices

yfinance · adjusted closes

Daily OHLCV for QQQ, last 52 weeks.

get_news

Polygon.io · news + sentiment

Recent headlines for QQQ + top 10 holdings, with per-article sentiment.

get_macro

FRED · point-in-time vintages

Fed funds, CPI, unemployment, 10-year yield, VIX.

get_earnings_calendar

Finnhub

Earnings events for the target week.

compute_indicators

in-house numpy

RSI(14), MACD(12,26,9), Bollinger Bands(20, 2σ).

submit_forecast

schema-validated

Final structured output: per-signal direction + reasoning + commentary.

How confidence is computed

The model never picks a confidence number itself. It chooses a direction for each of three independent signal categories (macro, technical, sentiment), and the platform computes confidence deterministically from the alignment.

Signal patternDirectionConfidence
All 3 agreeAligned direction75%
2 of 3 agree, weight-alignedMajority direction60%
2 of 3 agree, heaviest signal dissentsMajority direction55%
1 signal stands, others split / neutralSingle signal direction45%
Split decision (e.g. 1 bull / 1 bear / 1 neutral)Neutral50%

Default signal weights: macro 0.40, technical 0.35, sentiment 0.25. These are tuning parameters and may change post-launch.

Honest calibration

We backtested the agent on the trailing 0 weeks of QQQ history. Each historical Sunday, the agent was invoked with tools clipped to data available at that date — no future-knowledge leakage.

Forecasts scored

0

Strict win rate

Always-bullish baseline

55.8%

What this means. Over the last 0 weeks, naive "always predict bullish" would have been right 55.8% of the time. Beating that baseline on a 1-week QQQ horizon is genuinely hard — most professional fund managers don't. The directional lean we publish is one input among many; we recommend treating the per-signal breakdown and commentary as the primary value.

Weighted accuracy (giving partial credit for near-misses, e.g. a bullish call when the market closes flat) is . Both numbers are recomputed every Sunday after the previous week's close.

See full track record →

What this is not

  • Not investment advice. Educational research product. We don't know your portfolio, risk tolerance, or goals.
  • Not real-time. One briefing per week, generated Sunday night, valid until next Sunday.
  • Not a magic predictor. The directional call is wrong roughly as often as it is right. Read the commentary.
  • Not personalized. Same briefing for every subscriber. We don't take portfolio data or trade on your behalf.
  • Not financial advice. Consult a licensed advisor before making investment decisions.

Data sources

  • yfinance — Adjusted historical OHLCV for QQQ.
  • FRED (Federal Reserve Bank of St. Louis) — Macro series with point-in-time vintages so backtests don't cheat.
  • Polygon.io — News headlines, summaries, and per-article sentiment for QQQ + top 10 holdings.
  • Finnhub — Earnings calendar.
  • Anthropic Claude (via Amazon Bedrock) — The agent that synthesizes the above into the briefing.