PLTR news sentiment — contemporaneous vs next-day predictive power (last ~36 months)
Headlines and sentiment scores are often touted as a ready-made signal for tomorrow’s moves. To test that claim for PLTR, I paired ~36 months of daily mean news-sentiment with the same day’s and the next day’s returns (418 matched days), then measured Pearson correlations and simple OLS slopes to see whether sentiment leads returns or merely mirrors them.
The short answer: it mirrors. Sentiment vs same-day return shows a small but significant correlation (r = 0.163, p = 0.0008; OLS slope ≈ +0.039), while sentiment vs next-day return is indistinguishable from noise (r = 0.028, p = 0.572; slope ≈ +0.0067). The full analysis below presents the rolling statistics, tercile contrasts, and diagnostics that support that conclusion.
For PLTR over the past ~3 years, does the daily news-sentiment score actually lead the next day's return, or does it just mirror the same day's move? I expect sentiment to correlate tightly with the contemporaneous daily return but carry essentially zero predictive edge for next-day returns — the headlines chase the price rather than forecast it. This tests a popular, debunkable belief (that news sentiment is a tradeable leading signal) using a robust sample of daily sentiment scores against same-day vs. next-day returns.
How this was measured
Resampled PLTR minute bars to daily closes to compute same-day (ret_t) and next-day (ret_{t+1}) returns. Aggregated PLTR_news to a daily mean sentiment (ticker_sentiment_score preferred; overall_sentiment_score fallback). Joined price and sentiment on calendar day within the last 36 months. Measured Pearson correlations and fitted simple OLS (scipy.stats.linregress) for ret_t ~ sentiment_t and ret_{t+1} ~ sentiment_t. Reported 60-day rolling correlations and a tercile contrast on next-day returns.
The key numbers
Reading the numbers
Across 418 days, daily sentiment shows a small but statistically detectable contemporaneous link to same-day returns (r≈0.163, p≈0.0008), while the correlation with next-day returns is essentially zero (r≈0.028, p≈0.572) — no clear lead.
The charts
The 60-day rolling lines make the contrast obvious: the same-day correlation stays consistently positive (mean ≈0.1425, peaked at 0.4542 and ended near 0.359) while the next-day correlation hovers around zero (mean ≈0.0127) and crosses negative at times. Look at the gap between the two series — same-day is regularly above next-day across the window. That pattern supports the idea that sentiment moves with same-day returns rather than leading them.
The scatter of sentiment vs same-day return shows a positive slope but a lot of spread: sentiment ranges from -0.4451 to 0.8235 while same-day returns range from -0.1508 to 0.2148, and the OLS slope is about 0.039 with r≈0.163 (r²≈0.027). The association is statistically detectable (p≈0.0008) but explains only a few percent of return variation, so sentiment and contemporaneous price move together modestly rather than tightly predicting large moves.
The scatter for sentiment vs next-day return looks essentially cloud-like with no clear tilt: next-day returns span -0.1508 to 0.2451 while sentiment spans the same -0.4451 to 0.8235, and the regression slope is near zero (β≈0.0067) with r≈0.028 and p≈0.572. In plain terms, there is no detectable linear lead effect of today's sentiment on tomorrow's return in this sample.
Splitting days into terciles shows next-day means of roughly -0.0008 for Negative (n=140), 0.005 for Neutral (n=139), and 0.0033 for Positive (n=139) — the group means are tiny and very close. The boxplots overlap heavily and the Welch t between Positive and Negative is only about 0.85, indicating no clear tercile separation of next-day performance. This reinforces the conclusion that higher sentiment days do not produce a reliable next-day return edge.
Linear regression summary (returns ~ sentiment)
| model | beta | intercept | r | r2 | p_value | n |
|---|---|---|---|---|---|---|
| Same-day: ret_t ~ sent_t | 0.039438 | -0.003983 | 0.1631 | 0.0266 | 0.0008 | 418 |
| Next-day: ret_{t+1} ~ sent_t | 0.006707 | 0.001292 | 0.0277 | 0.0008 | 0.5723 | 418 |
Next-day return by sentiment tercile
| bucket | n | mean | std |
|---|---|---|---|
| Negative | 140 | -0.0008 | 0.0389 |
| Neutral | 139 | 0.005 | 0.0385 |
| Positive | 139 | 0.0033 | 0.0421 |
The takeaway
No — PLTR daily news-sentiment tracks the same day's move but offers no meaningful predictive edge for the next day. Over 418 matched days, sentiment vs same-day return shows r=0.163 (p=0.0008) with an OLS slope ≈+0.039 and r²≈0.027, while sentiment vs next-day return is r≈0.028 (p=0.572) with slope ≈+0.0067 and r²≈0.0008. The same-day relationship is statistically clear (about a 1-in-1,250 chance this is luck) but very small in magnitude — sentiment explains only ~2.7% of same-day return variance. The next-day result is indistinguishable from noise (p≈0.57), and tercile means (Positive 0.0033 vs Negative -0.0008; Welch p≈0.395) don't show a reliable lead. In short, headlines appear to mirror contemporaneous price action, not forecast it. If you need a next-day trading signal, this daily sentiment series over these 418 days does not provide one.
The fine print
- Timestamps are not intraday-aligned; end-of-day headlines can reflect moves already realized.
- Inner join excludes days without headlines, biasing the sample toward eventful/high-volatility days.
- Daily sentiment is an unweighted mean across headlines, so heavy mixed-tone days are averaged toward zero.
- Window covers ~36 months (2023-05-29 to 2026-05-29); results may differ in other regimes or longer histories.