← Back to dashboardTürkçe
Methodology · How We Measure

How a post moves the market — and why measuring it is hard

This page explains the methodology behind every score and chart on the site: the data we used, how we scored it, the mistakes we corrected, and how much confidence we can place in the current results.

1 · Data source

We obtained all of Trump's Truth Social posts since 2022—more than 35,000 posts—from CNN's mirror of the stiles/trump-truth-social-archive repository. The mirror updates every 5 minutes. For market data, we use two layers: SPY (S&P 500) for the broad market and Mag7 stocks (Apple, Microsoft, Nvidia, Alphabet, Amazon, Meta, Tesla) for company-level detail. We use daily and hourly prices from 2018 to the present. Section 3 explains which measure is appropriate and when.

2 · Scoring — how we grade a post

01
We query a wisdom engine using the post text. This local search system contains thousands of finance books and market lessons. We extract several of the most relevant lessons from two sources—book lessons and general market lessons—weighted equally.
02
The post text and selected lessons are fed into a language model. The model also receives the post date to reduce look-ahead bias. It is instructed: "This post was made on that date. Evaluate it using only the knowledge available on that date."
03
The model returns a score between -5 and +5 and a brief rationale in Turkish.
-5
strongly negative
-3
negative
0
neutral
+3
positive
+5
strongly positive

Example: a major trade shock such as a "100% tariff" receives a −5; a personal statement such as "America is a great country" receives a 0; a trade deal announcement might receive anywhere from +3 to +5.

3 · Market reaction — two questions, two measures

If a stock rose 2% after a post, did the post cause the move, or did the stock simply rise with the broader market? The standard approach is the abnormal return (AR) method. Subtract the index return from the stock return, and the remaining difference represents the movement specific to that stock:

Abnormal return

stock_return − QQQ_return
The primary measure used in event studies in the finance literature.

But there is a catch

This method is appropriate when a post concerns a specific company, such as Apple, Tesla, or a chipmaker. However, we found that the only topic through which Trump affects the market is tariffs. A tariff is a macro shock that affects the entire market at once. In that case, the broader-market movement being subtracted is the signal we are trying to measure. By definition, abnormal return washes out the macro shock.

We encountered this problem firsthand. The abnormal-return analysis found nothing. When we repeated the analysis using raw return, the tariff signal emerged.

That is why the site uses both measures. The main number on each card is the broad market's (SPY) raw return, which is the appropriate measure of a macro effect. Company-level Mag7 results use abnormal return and are meaningful only when a post concerns that specific company.

Why SPY rather than Mag7

We compared six instruments side by side. The effect of the tariff signal on the magnitude of the next day's movement was notably clearer in broad indices (t-statistic, 2025-2026): VTI +4.11 · SPY +4.02 · DIA +3.97 · QQQ +3.47 · IWM +3.27 · Mag7 median +3.28. Although Mag7 had the largest coefficient, it had the lowest t-statistic because a seven-stock median is noisy: its baseline volatility is 1.13%, compared with SPY's 0.51%. In short, the seven-stock measure added clutter while providing weaker statistical evidence.

The time window also matters. If a post was published in the evening or at night in New York, while the market was closed, the reference point is the next open. If it was published while the market was open, the reference point is the previous close. This avoids measuring the aftermath of the reaction rather than the reaction itself.

4 · What we found

The most counterintuitive finding

War rhetoric did not move the market. Military and war-related posts produced a result close to zero, with t=+0.02 in the 339-day sample. The results were t=-0.98 for Iran-Russia-Israel, +0.25 for China, +0.08 for the Fed, -0.01 for energy, and -1.25 for the courtroom topic. The most plausible interpretation is that Trump produces military rhetoric so often that the market has become desensitized to it.

Post scores also do not predict the market's direction the next day. The correlation between the score and the next day's return has remained near zero in all five years: 2022 +0.015; 2023 +0.015; 2024 +0.003; 2025 +0.005; 2026 -0.044. The sign can change from year to year, and the result is the same for both abnormal and raw returns. We therefore abandoned the idea of a bot that makes directional predictions from posts.

The only surviving signal: tariffs
When two or more tariff or trade posts are published in a single day, the magnitude of the market's movement increases the next day. The 2025-2026 coefficient is +0.596, t=+4.38. This signal does not indicate direction. It indicates only that realized volatility will increase.
But that signal is also fading
The tariff effect was +1.056 (t=3.53) in the first half of 2025, +0.171 (t=1.16) in the second half, and +0.242 (t=1.07) in 2026. The market appears to have learned to price in tariff threats.
The finding we retracted

Across the full dataset, we initially found “a 29% probability of a 2% or larger move following tariff days, compared with 15% on other days.” But in a walk-forward test using only the data available as of each day, the difference disappeared: 20.8% versus 20.1%. This was a classic in-sample illusion, so we do not present the initial result as a valid signal.

The only remaining asymmetry is in downside tail risk: the probability of a drop of more than 3% the next day is 8.3% following tariff days and 4.7% on other days. This difference is not sufficient for a directional prediction. It shows only that the risk of a large negative move is not evenly distributed.

In summary, across 27,564 posts, the general sentiment score does not produce directional predictions. War, geopolitics, and other prominent topics also fail to provide a reliable market signal. The only relationship that survives in the data is between intense tariff rhetoric and the magnitude of the next day's movement. The weakening of that relationship over time suggests that the market is adapting to repeated political threats.

5 · What we can and cannot rely on right now

The honest assessment

Most posts are unrelated to the market. Only about 7% of the 27,564 scored posts receive a non-neutral score—that is, a score other than 0—because most posts contain genuinely political or personal content. This is not a bug. It likely reflects reality.

No directional relationship has been established—in fact, the evidence indicates that none exists. The sample is no longer small. Across 27,564 posts, there is no reliable relationship between the score and market direction. None of the numbers on the site should be treated as a predictive tool. They are a record of past results.

No weight has been incorporated into a trading bot. Our internal rule is simple: an unproven statistic cannot receive a high-confidence weight. It must first be tested on a separate holdout set. The monitoring service for the tariff-volatility relationship continuously measures its own predictive power. Because the relationship is currently in a “signal faded” state, the service takes no action.

6 · Known limitations

Look-ahead bias
The language model may have been trained on data through 2025-2026. When scoring an older post, it might "remember" what happened afterward. Providing the date reduces this risk but does not eliminate it.
Multiple posts on the same day
Trump sometimes publishes dozens of posts in a single day. That day's market movement may become associated with all of them. We show only one post from the same day in our lists, but this does not fully solve the problem.
Hourly data coverage
The "immediate" reaction—the first hour after a post—is measured using IEX exchange data, which represents ~2-3% of total national volume rather than the full market. It is sufficient for analyzing the pattern and timing of a move, but not for full volume analysis.
Single-source scoring
Scores currently come from a single language model. We have not yet cross-validated them using multiple independent evaluators.

7 · So why does this site exist?

Because this is a starting point, not a final claim. The goal is to accumulate enough data, measure it correctly, and assess honestly whether a meaningful relationship exists. If one emerges, that matters. If not, that is also a valuable result: "social media noise does not move the market as much as people think" is itself a finding.