ECB similarity Research Paper Replication & Exentention

Terms :

CAR : Cumulative Abnormal or Excess returns of the MSCI Euro Index (EUR)

Findings & Summary

• Project replicates ECB Introductory Statement text-similarity (1999–2023), linking communication to market reactions.
• Methods: NLP (bigrams Jaccard, LM tone), statement-frequency ΔMRO, macro controls, HAC(6) errors, |CAR| over ±1/±3/±5/±7 days.
• Pre-2014: similarity increases with time (convergence), consistent with more templated language.
• Post-2014: similarity declines with time (greater novelty), robust to controls and standardization.
• Market impact: tone×similarity raises absolute CAR post-2014 (notably at ±7d); tone alone is weak.
• Controls: Output gap and inflation sometimes dampen |CAR| post-2014; ΔMRO levels/moves have limited incremental power.
• Bottom line: ECB communication style shifted after 2014, and tone + novelty together move markets more than tone by itself.

Figure 1 — Statement similarity over time

Figure 2 — The Pessimism measured using the Loughran-Mcdonald dictionnary.

Results (HAC(6), main |CAR| window ±5d, winsorized 1%)

Table 3 — Statement similarity vs time (Depvar: logSimilarity)

Sample	Spec	Key regressor	Coef	Sig	R²
1999–2023	(1) baseline	logTime_days	0.149		0.001
	(2) + controls	logTime_days	−0.122		0.047
	(4) count + controls	logTime_count	−0.113		0.047
1999–2013	(1) baseline	logTime_days	0.724	★	0.055
	(2) + controls	logTime_days	0.426		0.071
	(4) count + controls	logTime_count	0.488	★	0.072
2014–2023	(1) baseline	logTime_days	−8.489	★★	0.059
	(2) + controls	logTime_days	−15.526	★★	0.109
	(4) count + controls	logTime_count	−20.498	★★	0.104

Takeaway: Similarity rose pre-2014 (convergence) but fell post-2014 (greater novelty).

---

Table 4 — Market impact (Depvar: |CAR|_w, ±5d)

Sample	Spec	Regressor	Coef	Sig	R²
1999–2023	(1) baseline	pessimism	0.148		0.004
	(3) interaction only	pessimism×similarity	0.614		0.004
	(4) + controls	pessimism×similarity	0.760	★	0.034
1999–2013	(1) baseline	pessimism	0.176		0.004
	(3) interaction only	pessimism×similarity	1.604		0.004
	(4) + controls	pessimism×similarity	1.169		0.035
2014–2023	(1) baseline	pessimism	0.296	★	0.030
	(3) interaction only	pessimism×similarity	0.695	★★	0.020
	(4) + controls	pessimism×similarity	0.816	★★	0.058

Window robustness (post-2014): ±1d: n.s. • ±3d: interaction n.s., OutputGap − (p≈.01) • ±7d: interaction + (p≈.01), OutputGap −, Inflation −.

Legend: ★ p<0.10, ★★ p<0.05 (HAC SEs).

How we measure Similarity and Pessimism

Similarity (statement-to-previous statement)

• Scope: only the Introductory Statement segment (Q&A and boilerplate trimmed).

• Preprocess: lowercase → strip punctuation → remove English stop-words → Porter stemmer → build token list.

• Construct bigrams (pairs of consecutive tokens) and take set union per statement.

• Compute Jaccard similarity vs the immediately preceding meeting:

  $$ \text{sim}_t=\frac{|B_t \cap B_{t-1}|}{|B_t \cup B_{t-1}|} $$

• Use logSimilarity = log(max(sim_t, 1e-9)) to stabilize tails for OLS.

Pessimism (tone of the statement)

• Tokens: same text segment, but without stemming (keeps dictionary words intact).

• Dictionary: Loughran–McDonald finance lexicon (Positive/Negative lists).

• Counts: neg_t = #Negative words, pos_t = #Positive words, N_t = total tokens.

• Measure:

  $$ \text{pessimism}_t = \frac{\text{neg}_t - \text{pos}_t}{N_t} $$

• Used directly and in the interaction pessimism × similarity; standardized (z-scored) in some specs.

Reference: Amaya & Filbien (2015), Journal of Financial Markets. ScienceDirect link