Market sentiment

Problem

COS30018 Task C.7: predict stock price movement, and test whether news sentiment adds signal beyond technical indicators. The design is a clean ablation — a baseline classifier trained with sentiment features removed against an enhanced model with the full set — plus a separate regression pipeline forecasting actual prices. The honest question underneath: when the run is over, what can the code still prove?

Approach

News comes from Alpha Vantage’s NEWS_SENTIMENT feed (clients for Finnhub and NewsAPI were written but never wired into the pipeline — capability, not usage). Each article’s title and summary are scored three ways — VADER, TextBlob, and FinBERT — joined by the feed’s own score, then averaged per trading day into an ensemble sentiment with its disagreement spread. Every source-ticker-range fetch is cached to CSV, so re-runs skip the network entirely. Feature engineering produces 35 features in the configured run: OHLCV, eleven sentiment aggregates, seven technical indicators, nine lags, and three interaction terms — sentiment × 10-day volatility, sentiment × log-volume, and a sentiment-agreement score computed as the inverse of the tools’ disagreement.

Architecture

Two pipelines, deliberately separated — that separation is the refactor, receipted in the git history: three standalone scripts collapsed into one argparse CLI (collect-news → analyze-sentiment → train-models), the regression code split into its own package, 630 lines added against 1,330 deleted. Classification wraps logistic regression, random forest, and XGBoost behind one class with predict and predict_proba semantics; splits are chronological and the scaler fits on training data only. The regression side builds configurable LSTM/GRU stacks — default two GRU layers of 128 units reading 60-day windows to forecast five days — with a hybrid twist: a SARIMA model fit on the network’s training residuals, added back onto its forecast.

Results

This chapter is deliberately empty of accuracy numbers, and that is the finding. The pipeline demonstrably ran — 667 trading days of multi-tool sentiment for IBM and Coca-Cola sit in the caches with real FinBERT scores — and the ablation harness prints ΔF1 and ΔAUC against a 0.01 significance bar. But nothing saved them: every metric went to stdout, every plot to a window. The only numbers in the archive live in an AI-assisted analysis document that cites an experiment absent from the config — so under this site’s rules they are not results, and the accuracy card above stays an em-dash until the submitted report is retrieved. One methodological flaw is disclosed rather than hidden: the regression pipeline’s early stopping monitored the test set.

Reflection

The lesson is epistemic: code that ran is not the same as results you can cite. Two claims I would once have put on a resume died under verification — a “67% fewer API calls” figure that appears nowhere in the history, and an “sklearn-style interface” that trains with train(), not fit(). What survives is what the receipts support: cache-first collectors that skip repeat calls entirely, a real consolidation refactor, and a sound experimental design whose outputs were never persisted. The fix list writes itself: log every run to disk, wire the two dormant news sources, rerun the ablation, and backfill the pending card with a number that can defend itself.