Model
Tomorrow Signal (Local Gemma)
Hourly 'should I buy gas tomorrow?' signal. A locally-hosted Gemma 4 E4B (~8B-parameter MoE) reads the news, scores sentiment, and that score is blended with intraday RBOB gasoline futures into a 24-48 hour call. Nothing in the pipeline hits a cloud LLM — it all runs on our laptop.
What the model pays attention to
Approach
A cron job runs the pipeline once an hour. It pulls ~20 headlines from Google News RSS ('gas prices OR oil refinery OR OPEC OR crude oil'), dedupes by title, and sends each one to Gemma 4 E4B over Ollama with a short few-shot prompt that asks for UP / DOWN / UNCLEAR relative to pump-price pressure over the next 24-48h. The scores get averaged into a single sentiment number between -1 and +1. In parallel, the pipeline also pulls 5-minute RBOB gasoline futures (RB=F) from Yahoo Finance and computes 1h / 6h / 24h / 5d returns and a 5-day z-score. Both signals go into a 60/40 weighted blend (sentiment / price), through a sigmoid, thresholded at 0.5. The result is a tomorrow YES or NO with a confidence number. The whole sentiment side of this runs locally — no OpenAI, no Anthropic, no Google API key in the loop.
Features
Built by
Team lead on the project. Built the intraday 'should I buy gas tomorrow?' pipeline after class feedback asked for a shorter-horizon signal.
Trained on
Not trained — rules-based combiner. 10/10 on an internal headline-labelling benchmark.