A daily research scan of a small watchlist. Flags mean-reversion setups — names that have pulled back into a dip pattern — and confirms or rejects them with FinBERT sentiment on the last 5 days of headlines. Reruns every weekday at 17:30 ET after US close.
WATCH — dip rule fires: close < SMA50×0.95 and RSI < 35 and 20d drawdown > 5%.
CONSIDER — WATCH plus sentiment confirms: mean sentiment < −0.3 across ≥3 confident headlines.
— — quiet, no dip today. Most days look like this. That's the expected state.
| Ticker | Close | RSI(14) | 20d DD | SMA50 gap | Headlines | Sentiment | Signal |
|---|---|---|---|---|---|---|---|
| AAPL | $310.85 | 79.1 | 0.0% | 14.0% | 177 / 205 | -0.03 | — |
| MSFT | $412.67 | 50.1 | -2.6% | 3.1% | 200 / 244 | -0.01 | — |
| NVDA | $212.60 | 51.0 | -9.8% | 7.3% | 218 / 248 | 0.04 | — |
| GOOGL | $388.83 | 61.0 | -3.4% | 12.9% | 208 / 249 | 0.02 | — |
| AMZN | $271.85 | 62.7 | -1.1% | 11.3% | 213 / 245 | 0.05 | — |
| META | $635.26 | 56.9 | -5.1% | 2.8% | 189 / 224 | -0.05 | — |
| TSLA | $440.36 | 63.0 | -1.1% | 12.9% | 179 / 208 | 0.07 | — |
| SPY | $750.46 | 71.2 | -0.0% | 7.2% | 194 / 224 | 0.03 | — |
| QQQ | $729.45 | 74.6 | -0.1% | 12.7% | 81 / 91 | 0.03 | — |
This is a contrarian filter. Numbers that look ugly on a normal stock screen are exactly what we want to find — oversold RSI, deep drawdown, broken trend, negative news flow. The thesis: when those four red flags align on a fundamentally sound stock, the market is being too emotional and reversion is likely.
The trap: bad-looking is good only when the dip is temporary. When fundamentals genuinely broke, the same bad-looking numbers mean "catch a falling knife." The dashboard cannot distinguish these — that's why a CONSIDER is a prompt for due diligence, not a buy trigger.
The deeper the numbers go past my thresholds, the more carefully you need to verify why before acting. The rule is calibrated for noise-driven dips in mostly-healthy megacaps; pathological situations land in the same bucket but mean the opposite thing.
The 14-day Relative Strength Index measures the speed and magnitude of recent price changes. Computed using Wilder's smoothing of up-day gains vs down-day losses over the trailing 14 trading days.
| Range | What it means | Dip-buy view |
|---|---|---|
| 70+ | Overbought, ran too far too fast | ❌ No dip here |
| 50–70 | Normal uptrend, mild pullback | 🟡 Nothing yet |
| 35–50 | Weakening | 🟡 Watch but not signal |
| < 35 | Oversold — my threshold | ✅ Contributes to dip signal |
| < 20 | Extreme oversold / capitulation | 🟢 Strongest or big trouble |
Limitation: RSI can stay below 30 for weeks in a real bear move. "Oversold" doesn't mean "guaranteed to bounce." Combine with the other columns.
Measures how far the current close is below the highest closing price of the last 20 trading days (about one trading month). Always ≤ 0% by construction.
DD = (Close / max(Close over last 20 trading days) − 1) × 100
Why it matters: direct measure of "how much has this stock fallen from its recent peak" — the literal definition of the dip you'd be buying. Trend-agnostic: works in uptrends and downtrends alike.
| Range | Eyeball as |
|---|---|
| 0 to −2% | At recent highs — no pullback |
| −2 to −5% | Normal chop — every stock wobbles this much |
| −5 to −10% | Real pullback — my threshold (DD > 5%) fires here |
| −10 to −15% | Sharp correction — pain showing |
| Worse than −15% | Crash territory — something is happening |
Limitations: the 20-day window resets when a new high prints — a stock that fell 30% then bounced 20% might show only −10% DD even though it's near year-lows. DD also says nothing about why the stock fell.
Distance between the current close and the average of the last 50 closing prices (~2.5 months of trading). Positive = above trend. Negative = below trend.
gap = (Close / SMA50 − 1) × 100
Why it matters: the 50-day moving average is the most-watched intermediate-term trend line in equity markets. Every chart, every algo references it. When a stock breaks below SMA50, trend-followers exit and "trend is broken" gets written into research notes. It's self-fulfilling — enough capital treats SMA50 as the line that it becomes the line.
For dip-buying: a stock that's just dipped (DD shows recent pain) AND broken below SMA50 (trend-followers have bailed) is in maximum forced-selling mode — exactly the irrational-pessimism state contrarians want.
| Gap | Eyeball as |
|---|---|
| > +10% | Stretched — momentum chase |
| +5% to +10% | Strong uptrend, healthy |
| 0 to +5% | At trend — normal |
| 0 to −5% | Trend wobbling but intact |
| < −5% | Trend broken — my threshold (Close < SMA50 × 0.95) fires |
| Worse than −10% | Major breakdown — likely something fundamentally wrong |
Limitations: choppy stocks can cross SMA50 daily on noise. Threshold magnitude (−5%) suits megacap tech; would be too loose for crypto, too tight for utilities.
FinBERT (a BERT variant fine-tuned on financial text) scores each Finnhub headline as positive / negative / neutral. We keep only headlines where FinBERT was confident (its max-class probability > 0.6) and average positive − negative across them. Range: −1 (all bearish) to +1 (all bullish).
| Range | News flow looks like | Dip-buy view |
|---|---|---|
| +0.3 to +1.0 | Bullish (earnings beats, upgrades, hype) | ❌ Already optimistic |
| +0.1 to +0.3 | Mildly positive | ❌ Boring |
| −0.1 to +0.1 | Mixed / neutral — normal day | 🟡 Most days are here |
| −0.1 to −0.3 | Mildly negative | 🟡 Early concern |
| −0.3 to −0.6 | Clearly negative — "fear" cycle | ✅ My CONSIDER threshold |
| Worse than −0.6 | Extreme negative | 🟢 Capitulation or something legitimately broken |
Headlines column matters: n_confident / n_total. If confident count is < 3, the sentiment number is essentially undefined and CONSIDER won't fire regardless. Healthy reading is ~85–95% confident.
Limitation: FinBERT is trained on long-form financial text (~2019). It handles headlines but misses some modern tone. A v2 Claude-Haiku swap is parked in our roadmap.
The metrics measure different things. The rule requires both a recent drawdown AND a broken longer-term trend — and the news layer further filters from there:
SMA50 gap
Above trend Below trend (< −5%)
───────────── ──────────────────
DD shallow │ ❌ Strong │ ⚠️ Slow grinder
(0 to −5%) │ uptrend │ No recent peak —
│ No dip │ stock has been
│ │ bleeding quietly
│ │ (no signal fires)
─────────────────────────────────────────
DD deep │ ⚠️ Spike-and- │ ✅ CLASSIC DIP
(< −5%) │ revert │ Recent sharp drop
│ Recent sharp │ PLUS trend broken —
│ drop, trend │ short-term sellers
│ still intact │ AND trend-followers
│ │ both bailing
│ (no signal) │
│ │ → WATCH or CONSIDER
Top-right (slow grinder): real decline but no "dip" to buy — nothing to revert from. Catch-a-knife territory.
Bottom-left (spike-and-revert): flash sell-off in a strong uptrend. Often mean-reverts on its own but the bigger thesis ("everyone's bailing") doesn't apply because nobody actually bailed yet.
Bottom-right: where the rule fires. Forced selling + trend-follower exits = maximum pessimism = the actual mean-reversion edge.
Layer in the sentiment column on top: if news flow is negative at the same time, the bet upgrades from WATCH to CONSIDER — meaning the pessimism has a story attached to it, which is contrarian-bullish provided the story is not real damage.
Do nothing. Most days look like this and that's the expected state. The tool is silent until conditions align.