ACG 6805 portfolio project · live experiment
When Reddit suddenly talks about a stock, does the price follow?
Our bot watches hourly Reddit mention counts, scores the news, and paper trades the biggest spikes. Every decision is logged, traded or skipped, and the results are compared against simply buying the market.
Loading…
Portfolio vs. the market
All three start at 100 on day one, so the lines show percent change side by side.
Benchmarks are buy-and-hold of SPY (S&P 500) and IWM (Russell 2000) over the same dates. Slippage applies a 2% cost to each buy and each sell, because real fills on fast-moving stocks are usually worse than paper fills.
Trade log
Newest first. $1,000 per buy; sells on a 25% stop-loss, 30% gain, or after 3 trading days.
| Ticker | Bought | Trigger | Exit | P&L |
|---|
Signal funnel
Every spike is logged, including the ones we skipped, so the test isn't just our winners.
Why spikes were skipped
How it works
Six small Python scripts run on a schedule and share one database. No AI model makes trading decisions; every rule is written down in a settings file.
- 1
Scout
Every hour, saves how often each ticker is mentioned on Reddit and 4chan (via ApeWisdom), and how that compares with 24 hours earlier.
- 2
Sentiment
For the biggest spikes, scores the last 48 hours of news with VADER, a rule-based tone scorer: Bullish, Neutral or Bearish.
- 3
Filter
Keeps stocks priced $0.50–$1,000 with at least $1M traded a day, mentions at least doubled, and no bearish news. Logs a reason for every skip.
- 4
Trader
Ranks what passed and buys up to 8 per run, $1,000 each, with limit orders in an Alpaca paper account.
- 5
Exits
Sells at a 25% loss, a 30% gain, or after 3 trading sessions, whichever comes first.
- 6
Report
Summarizes the day and feeds this page, comparing results with the S&P 500 and Russell 2000.
What we changed after testing
Each fix came from watching the bot run, finding where it behaved differently than intended, and adding a test so it can't come back.
Runs arrived late or not at all
GitHub's built-in scheduler skipped hourly runs and started the trader after market close. We moved scheduling to an outside timer that triggers each run on time, in Eastern time.
Weekends counted as holding days
Stocks bought Friday were sold Monday after about 6.5 hours of trading. Holding time now counts trading sessions, skipping weekends and holidays.
Sold and re-bought the same stocks in one day
The bot paid two spreads to end up holding the same four stocks. A two-session cooldown now blocks quick re-buys.
One bad news article could hide a stock's real tone
A review found a single malformed article wiped out a stock's sentiment, letting bearish stocks through. Bad articles are now skipped individually.
Evening reports showed "no data"
The report used UTC, so after 7 PM Eastern "today" became tomorrow. It now uses the market's own day.
Governance and limitations
Data sources
Mention counts from ApeWisdom; news and prices from Alpaca (Benzinga news). No personal data is collected. News headlines are scored, not republished.
Where AI was used
Claude helped write and review the code. The bot itself uses no AI model: VADER is a fixed word list, and every trading rule is set in a config file.
Human review points
We review every code change before it runs, check daily reports, and can pause the schedule at any time.
Limitations
Paper fills are optimistic. A few weeks of data can't separate skill from luck. Reddit buzz can be manipulated, which is part of what we're measuring.
Do not use this for
Real-money trading decisions, investment advice, or anything treated as a stock tip. This is a research experiment with simulated money.
About the project
Built by a two-person team for ACG 6805, Fall 2026.
Co-built an automated paper-trading system in Python that tests whether sudden Reddit mention spikes predict stock price moves, logging every signal and benchmarking results against the S&P 500 on a live dashboard.