Video summary
I Built a Bot to Trade Based on Reddit Posts
Main summary
Key takeaways
Finance-focused summary (markets/investing/portfolio/risk)
- The creator attempts to build an automated trading system using WallStreetBets (WSB) Reddit posts to predict when stocks will move “before they take off.”
- The core research question is whether “wisdom of crowds” from WSB can beat the market, using the S&P 500 as the benchmark.
- Multiple trading signals are tested, and all fail to outperform the index. Some approaches also produce very large drawdowns.
Instruments / tickers / assets mentioned
- S&P 500: baseline index for comparison
- GameStop: explicitly referenced, linked to the short squeeze dynamic
(No other specific tickers are explicitly named in the subtitles.)
Methodology / step-by-step frameworks tested
1) Naive “buy what’s mentioned”
- Scrape WSB posts in real time
- Scan each post for stock tickers
- Count mentions per ticker each day
- Run a randomized baseline:
- Monte Carlo: randomly select thousands of posts
- Buy the stock mentioned in each selected post
- Hold for 60 days, then sell
- Compare returns to buying the S&P 500 over the same day/period
- Key feature: no sentiment selection (posts may be bullish or bearish)
2) “Herd-following” via most-mentioned
- For a daily window (2021 to 2022):
- Count ticker mentions daily
- Identify the single most mentioned stock each day
- Buy it and hold until it stops being the most mentioned
- Rationale: people follow attention/hype
3) “Velocity” / acceleration of mentions
- Compute mention growth rate (“velocity”):
- Example: 5 mentions Monday → 10 Tuesday → 15 Wednesday
- Velocity defined as increase in mentions per day (incremental mentions)
- Daily routine:
- Rank tickers by velocity each day
- Buy the fastest riser
- Hold and then sell when the signal changes (described as “buy fastest riser then hold and sell”)
- Outcome: reported to be poor versus the index
4) Normalized velocity to remove subreddit growth
- Problem addressed: raw mention increases may reflect WSB growing overall
- Fix:
- Count ticker mentions as a fraction of total posts that day
- Equivalent to dividing by a 7-day rolling average of total posts (per a referenced Oxford paper)
- Still fails to beat the market
5) Text sentiment classification (Oxford approach, referenced)
- Oxford researchers:
- Scraped WSB posts 2012–2022
- Used AI to classify each post as bullish vs bearish
- Simulated trading and evaluated predictive power
- Manual labeling:
- 4,000 posts manually classified to train the model
Key numbers (performance metrics, timelines, explicit results)
Monte Carlo / random post picks
- Median return (WSB random picks): -10.3% over ~2 months
- Median return (S&P 500): +4.6% over ~2 months
Most-mentioned strategy
- Time window: 2021 to 2022
- Performance described qualitatively as driven by GameStop short squeeze
- Fails because the stock stays top-mentioned even after peak popularity → profits largely revert
Velocity strategy
- Trades executed: 222 trades over ~1 year
- Average strategy return: -1.1%
- Average S&P 500 return over the same period: +30%
- Best trade: +58%
- Worst trade: -83%
- Interpretation implied by results: severe volatility and poor risk-adjusted outcomes (very large tail losses)
Normalized velocity strategy
- S&P 500 return over the time window: +30%
- New strategy return: -0.2% over ~1 year
- Note: described as “better than before,” but still far below the index
Oxford sentiment model (referenced results)
- Sentiment prediction accuracy: 69%
- Trading simulation using sentiment:
- General case next-day loss: -2.8%
- During “GameStop year”: -4.4%
- Nuance: when isolating “due diligence” / researched writeups, returns reportedly flipped positive, though the subtitles don’t provide the exact magnitude for that subset
Recommendations / cautions explicitly implied
- The creator concludes that:
- Mention count / velocity / hype-based signals do not beat the index.
- Attention metrics are confounded by factors like overall subreddit growth and “front-page/news” effects (i.e., by the time it’s most mentioned, everyone already knows).
- Risk reality check from results:
- Even when wins occur (e.g., +58%), losses can be extreme (-83%), suggesting inadequate drawdown control.
Disclosures / disclaimers
- No explicit “not financial advice” wording appears in the provided subtitles.
- Results are framed as backtests/experiments, not guaranteed investing success.
Presenters / sources mentioned
- Oxford University researchers: referenced paper; scraping 2012–2022 and manual labeling of 4,000 posts
- Video creator / narrator: implied by statements like “I built… I ran… I post…” (name not provided in subtitles)
- S&P 500: benchmark/index (not a presenter)
- WallStreetBets (WSB): data source (not a presenter)