Abstract: Political parties and their members regularly hold speeches in which they express their opinions. Thereby, their speeches can reveal their underlying positioning – potentially in very high frequency. Herein, we propose a method for estimating high-frequency time-series of party positioning from their members’ parliamentary speeches. Our approach leverages two recent methodological innovations: pre-trained language models and dynamic scaling. We apply our approach to data covering parliamentary speeches within the German Bundestag during a 12-year period. Our monthly positioning estimates are highly consistent with a plethora of established benchmarks derived from manifesto texts, expert-surveys, roll-cast votes, or party-embeddings. In contrast to extant approaches, our estimates uncover substantial positioning dynamics across and within legislative periods. We demonstrate that simple measures of positioning dynamics can help to explain up to 20%p of additional variance in weekly election polls. [Show abstract ...] .collapsed { height: 5px !important; overflow: hidden; } .article-style { pointer-events: none; cursor: default; } .show_more { color: #1c3058; pointer-events: auto !important; } function wp_III_toggle_abstract() { var abstract = document.getElementById('wp_III_abstract'); if( abstract.className == 'collapsed' ) { abstract.className = 'expanded'; var max_height = document.getElementById('wp_III_abstract_text').clientHeight + 25; abstract.style = 'height: ' + max_height + 'px'; document.getElementById( 'wp_III_more' ).textContent = '[Hide ...]'; } else { abstract.className = 'collapsed'; document.getElementById( 'wp_III_more' ).textContent = '[Show abstract ...]'; }; }