NEXUS

Prior to this project, my finance knowledge could be generously described as "why yes, the stock market exists." Consequently, I spent the first week or so doing scant nothin' but eye-perusing: academic papers on momentum signals, behavioral finance research by Shiller and Thaler, the original Wilder (1978) paper introducing RSI, Bollinger's work on adaptive volatility envelopes, and a substantial amount of Federal Reserve documentation on yield curve dynamics.

A few concepts quite surprised me. The first was that quants never work with raw prices. They work with returns, which are the percentage change from one period to the next. The reason is stationarity: raw prices trend upward over time, meaning their statistical properties change constantly. Returns do not. This seems obvious in retrospect but is never mentioned in casual financial media.

The second surprise was the yield curve inversion signal. The spread between 2-year and 10-year U.S. Treasury yields has inverted before every single U.S. recession since 1955, without exception. The mechanism is elegant: banks borrow short-term and lend long-term. When short rates exceed long rates, that spread goes negative, lending becomes unprofitable, banks pull back on credit, and the economy contracts. This signal is freely available from the Federal Reserve's public API and is essentially invisible to most retail investors.

The third was the Sharpe ratio, which reframes strategy evaluation entirely. A strategy that returns 80% but swings wildly may be worse than one that returns 25% smoothly. Risk-adjusted thinking is the foundation of professional portfolio management, and it is almost completely absent from retail financial discourse.

I also spent time studying the architecture of existing tools: Quiver Quantitative (congressional trading tracker, alternative data), SentimenTrader (sentiment-focused with 2800+ indicators), Bloomberg Terminal (the institutional standard at $24,000 per year). The gap I identified was that no open-source tool combines technical, macro, sentiment, and cross-asset signals with interpretable plain-English output, built by someone who had to learn all the underlying math from scratch.

Before writing a single line of code, I designed the full data pipeline. The system is organized into four layers.

The Data Ingestion Layer lives in src/data/ and has three modules. finnhub.js handles real-time stock quotes, one year of daily price history, company news with sentiment scores, and company profiles (60 API calls per minute, free). fred.js pulls six macroeconomic series from the Federal Reserve: VIX, 10-year Treasury yield, 2-year Treasury yield, Federal Funds Rate, CPI, and unemployment rate. FRED is a public service with unlimited free requests. feargreed.js retrieves the CNN Fear and Greed Index via Alternative.me and Bitcoin/Ethereum prices via CoinGecko, both requiring no API key at all.

The Analysis Layer in src/analysis/ has four modules. technical.js implements RSI with Wilder smoothing, MACD with configurable EMA periods, Bollinger Bands with rolling standard deviation, rolling Z-scores, annualized volatility, ATR, Sharpe ratio, Sortino ratio, and maximum drawdown, all from scratch with no financial libraries. anomaly.js detects eight anomaly types including return outliers (|z| > 2.0), volatility regime shifts, RSI extremes, Bollinger squeeze detection, Bollinger Band breaches, golden/death crosses, volume spikes, and MACD crossover events. macro.js classifies the macroeconomic regime using a four-quadrant framework. backtest.js simulates four trading strategies on historical data and outputs Sharpe ratio, max drawdown, total return, and alpha versus buy-and-hold.

The Fusion Engine in src/engine/signal.js is the novel core of the system. It combines all four domains: MIS = (Technical x 0.40) + (Macro x 0.30) + (Sentiment x 0.20) + (Cross-Asset x 0.10)

The Display Layer in src/display/terminal.js renders the full dashboard using chalk for colors and cli-table3 for formatted tables, with spinner animations and color-coded signal outputs.

This was the phase I found most intellectually rewarding, and most humbling. Implementing every formula from scratch forced genuine understanding.

The RSI implementation revealed a subtle trap. The Wikipedia formula looks straightforward. The correct implementation uses Wilder's smoothing method, a modified exponential moving average, rather than a simple rolling mean. The two approaches produce noticeably different results. I verified my implementation against known RSI values from multiple sources before trusting it. Lesson: in quantitative finance, implementation details matter enormously. Two people implementing RSI can get results differing by several points.

The Z-score anomaly detection required careful handling of the window parameter. Too short (5 to 10 days) produces too many false positives. Too long (60+ days) smooths away genuine regime shifts. Twenty days emerged as a reasonable default for return Z-scores, following conventions in the academic literature.

The backtesting was the most humbling part. The RSI mean-reversion strategy, which seemed obviously correct in theory, dramatically underperforms buy-and-hold on many tickers in strong trending markets, because "oversold" simply means the trend is accelerating downward. The hybrid strategy outperforms because dual confirmation reduces false signals. But even the hybrid's parameters were chosen by looking at historical data, which is a form of overfitting. This is the central methodological challenge in quantitative finance, and acknowledging it openly is more intellectually honest than pretending a backtest is a performance guarantee.

Running NEXUS on live data for the first time was invigorating. A sample AAPL analysis on March 25, 2026 produced a MIS of 50/100 (Neutral/Mixed), with a technical score of 39 (RSI at 66.6 approaching overbought, MACD histogram just turned negative) fighting against a sentiment score of 68 (Fear and Greed at 14, Extreme Fear, which NEXUS interprets as a contrarian bullish signal). The two domains pulling in opposite directions produced a neutral composite, which is itself useful information.

The macro regime read Neutral/Mixed: VIX at 26.9 (elevated market anxiety), yield curve positive at +0.51% (not inverted), Fed Funds at 3.64% (on hold), CPI elevated at approximately 3.2% annualized.

The anomaly engine flagged a real signal: Apple's price was breaking above its upper Bollinger Band, statistically extended, flagged at MEDIUM severity. The backtest found the hybrid strategy best at +7.69% return versus +1.43% buy-and-hold, Sharpe ratio 0.842, max drawdown of only 6.29%. The Fear and Greed reading of 14 was particularly striking because it was a live real-time reading of genuine market anxiety, not demo data.

The Finnhub free tier restricts access to the historical candle endpoint, which I discovered only after the system was fully built. The fix was a graceful fallback: use synthetic price history for technical indicators while using real live data for everything else (quotes, news, FRED macro, Fear and Greed, crypto). The output is still genuinely useful and clearly labeled.

The most important intellectual lesson was about overfitting and the gap between backtests and reality. A backtest that looks good is not evidence a strategy works in real markets. Transaction costs, slippage, bid-ask spreads, and the fundamental fact that you chose parameters by looking at historical data all erode theoretical performance. NEXUS acknowledges this explicitly in its output. Any system that does not is either naive or dishonest.