Polish MPC minutes — lexical evolution
NBP MPC decision-making meeting minutes since April 2007. The tab reads each published minute, scores its language on a hawkish-dovish lexicon and eight macro topics, ranks each year's distinctive vocabulary via TF-IDF, and maps how Polish monetary-policy language shifted from cycle to cycle. — meetings indexed; the most recent meeting in the corpus is —.
How to read this page (methodology and glossary)
This page reads the published English-language minutes of every NBP MPC decision-making meeting from April 2007 onward. The text is scored on four layers: a hand-curated hawkish-vs-dovish lexicon tuned for Polish central-bank language, a TF-IDF model over unigrams plus bigrams with sklearn's English stopword list extended by roughly 130 central-bank ceremonial words ("noted", "underlined", "members", month names, etc.), a fixed list of eight macro topics tracked by direct keyword count, and pairwise cosine similarity across the meeting corpus.
The five tiles above
| Latest meeting tone | Net tone = (hawkish − dovish) / (hawkish + dovish) in the custom MPC lexicon, scored on the most recent published minute. Colored left border: salmon hawk-tilted, green dove-tilted, grey neutral. |
| Tone, 3m MA | Same metric averaged over the last three meetings. Smooths out single-meeting noise. |
| Last NBP move | Most recent reference-rate change in basis points, with the resulting rate level. |
| Trending term | The TF-IDF term that surged most in the last four meetings versus the prior baseline. |
| Policy regime | Mechanical classification of the trailing-6-month rate change: ≥ +25 bp = tightening (salmon), ≤ −25 bp = accommodation (green), otherwise neutral. |
What each panel shows
Hawkish-vs-dovish net tone with rate path. The lexicon score per meeting (light line) and as a 3-meeting moving average (heavy line). Background regime bands derived from the NBP rate path. The navy staircase on the right axis is the reference rate itself; triangles mark each policy action, labelled in bp for moves of 50 bp or more.
Eight macro topics. Per-1,000-word counts at each meeting for inflation, growth, labour, fiscal, energy, FX, geopolitics, and global, plotted as a 4×2 grid of small multiples so the relative magnitudes are directly comparable.
Lexical evolution time-lapse. The video at the top of the page sweeps through every MPC minute at one frame per meeting, with the trailing-six-meeting TF-IDF cloud at the top and the per-meeting net-tone strip below. Same vocabulary the page scores, animated rather than sliced by year.
One word per year. The single most distinctive TF-IDF term for each year, colored by the dominant monetary-policy regime in that year.
Term tracker. Type one or more words, comma-separated. Each gets its own line. Lookup is exhaustive — terms not pre-indexed are counted live from the meeting text, so any word that appears even once will plot.
Pairwise meeting similarity. TF-IDF cosine similarity between every pair of (monthly-sampled) meetings. Bright off-diagonal blocks mark periods where the MPC's vocabulary clustered around a single set of themes — GFC, deep cuts, COVID, 2022 inflation, the 2024-2026 pause.
Hawkish-vs-dovish net tone, with monetary-policy regime bands and NBP rate path
Net tone = (hawkish − dovish) / (hawkish + dovish) in a custom MPC lexicon, per meeting (light line) and as 3-meeting moving average (heavy line). Background bands derive mechanically from the 6-month change in the NBP reference rate. The navy staircase on the right axis is the NBP reference rate; triangles mark each policy action, labelled in bp for moves of 50 bp or more.
Eight macro topics, frequency over time
Per-1,000-word counts at each meeting for eight pre-defined themes, plotted as a 4×2 grid of small multiples so the relative magnitudes are directly comparable.
Lexical evolution time-lapse — a 1m46s video sweeps through every MPC minute since 2007, with a trailing 6-meeting word cloud and the net-tone playhead. Same vocabulary this page scores, animated.
▶ Watch on YouTube ↗ All movies →Track any word or bigram
Type one or more words, comma-separated. Each gets its own line. Lookup is exhaustive across the corpus — terms not pre-indexed are counted live from the meeting text, so any word that appears even once will work.
Lexical eras — Ward clustering on the pairwise-similarity matrix
Hierarchical Ward clustering on the cosine-distance matrix, cut at K clusters. Each meeting is assigned to one cluster. The strip below shows every meeting in chronological order, coloured by its cluster. The cards below the strip show the distinctive vocabulary of each cluster as a mini word cloud — these are the natural lexical eras the MPC's vocabulary fell into. Use the K-pill to vary the number of clusters; the underlying pairwise-similarity heatmap is available as a reveal at the bottom of this panel.
Pairwise similarity heatmap (212 × 212)
TF-IDF cosine similarity between every pair of meetings. Chronological orders rows and columns by meeting date — bright diagonal bands show within-era similarity. Clustered (Ward) reorders by Ward leaf order so the K cluster blocks line up as crisp diagonal squares.
Source-by-source observation timestamps for every input feeding this page are listed in the Data lineage & freshness block on the home page.