All MicroEvals
这是我的课程研究摘要和实验设计部分,你觉得投nature期刊的话有什么需要修改的地方吗?Abstract Do chan...
Create MicroEval

这是我的课程研究摘要和实验设计部分,你觉得投nature期刊的话有什么需要修改的地方吗?Abstract Do chan...

Prompt

这是我的课程研究摘要和实验设计部分,你觉得投nature期刊的话有什么需要修改的地方吗?Abstract Do changes in central bank wording improve risk forecasts once the announcement has occurred? We combine scheduled Federal Open Market Committee statements, public intraday announcement-window market changes and daily prices for eight exchange-traded funds. Statement revision is measured by the cosine distance between consecutive scheduled statements. A positive conditional second-moment model directly minimizes a regularized QLIKE criterion. A common penalty is selected using 2012–2014 data, and expanding quarterly estimation generates historical forecasts for 2015–2025. The principal comparison adds wording revision to a model that already includes market predictors, release and document-length controls, and absolute announcement-window Treasury and equity responses. Across 2,761 five-day forecast origins and 87 scheduled announcements, the mean QLIKE gain is 0.000919, with a 95% block-bootstrap interval of [−0.001303, 0.002732]. The gain is not statistically distinguishable from zero. It is also sensitive to a small number of influential dates. Alternative text representations, information delays, decay rates and announcement windows do not establish a reliable gain. Although the augmented model outperforms a simple exponentially weighted benchmark, a stronger market-only model has lower average loss. The evidence therefore supports a limited conclusion: this transparent wording-revision measure does not demonstrate incremental forecasting value under the examined design. 3 Data and research design 3.1 Public sources and coverage Table 1 Observed data coverage Source Coverage used Purpose Federal Reserve Board 167 scheduled statements, 2005–2025 Consecutive wording revisions Yahoo Finance 5,008 common ETF dates, 6 Feb 2006–31 Dec 2025 Adjusted-close return targets Cboe VIX aligned to the common trading calendar Previous-day market state SF Fed USMPD Scheduled statement windows and timestamps Immediate market-response controls The fixed asset universe is SPY, QQQ, TLT, GLD, EEM, VNQ, DBC and EFA. These funds cover distinct traded exposures, although several equity funds remain correlated. The universe is chosen for interpretable exposure and a common price history, not from an ex post ranking of returns. The intersection begins on 6 February 2006. Calendar 2006 supplies initialization; model estimation begins in 2007. 3.2 Statement extraction and event matching We archive official HTML pages and retain only events classified as scheduled statements by USMPD. The initial archive contains 177 candidate documents. Matching and publication-date checks retain 167 scheduled statements, including eight in 2005 for text-history initialization. The common price period contains 158 scheduled releases; the 2015–2025 evaluation contains 87. Emergency and other unscheduled releases are excluded from the text signal. Their realized market outcomes remain in daily targets and market predictors. The extractor removes voting paragraphs, related discount-rate actions, contact information and navigation. Displayed publication dates take precedence over digits embedded in a URL. This matters for the June 2007 statement: its archived address contains 20070618, whereas the page and event database identify 28 June 2007. Using the address as a date would make the document available ten calendar days too early. The replication archive records this correction and every inclusion decision. Each statement is compared with the previous scheduled statement, including when an unscheduled statement occurred between them. Accordingly, the measure represents scheduled-to-scheduled wording revision. It is not the difference from every intervening policy communication. The distinction is especially relevant around March 2020 and motivates separate reporting outside calendar 2020. 3.3 A transparent wording revision measure We lowercase alphabetic tokens, remove a fixed list of function words, and count the remaining unigrams. Negations are retained. There is no stemming, external sentiment dictionary, fitted embedding or full-corpus inverse-document-frequency estimate. Let cₑ denote the count vector of statement e. The revision measure nₑ is one minus the cosine similarity of cₑ and the preceding scheduled count vector. The union of the two documents supplies the dimensions; future documents have no role in the calculation. nₑ = 1 - (cₑ · cₑ₋₁) / (‖cₑ‖ ‖cₑ₋₁‖) Figure 1. Wording revision across scheduled statements. The shaded interval marks 2012–2014 validation; the dashed line marks the start of historical evaluation. Distance measures changed word use, not risk severity. 3.4 Daily information sets and forecast targets A release contributes its value on the announcement trading date and for nineteen subsequent trading dates. At age k, its weight is exp[−log(2)k/5]; values are summed without a changing denominator. The same twenty-day window and five-day half-life apply to occurrence, length, length revision, market responses and wording revision. This common transformation prevents a difference in persistence rules from being mistaken for incremental text content. Alternative half-lives are reported as sensitivity checks. Market controls B contain log squared current return, log means of five and twenty-two squared returns, the magnitude of the negative current return, and log VIX from the previous trading date. A fixed 10⁻¹⁰ floor is used only before taking logs of the return-based predictors. Observed outcomes are never floored. Using previous-day VIX avoids relying on a same-day index close whose publication time may differ from the ETF close. Table 2 Nested information sets Model Inputs Interpretation EWMA Squared returns with fixed decay 0.94 Simple external benchmark B Five market predictors Market-only conditional model C B plus release, log length, absolute log length change and two market-response kernels Matched communication control N C plus the wording-revision kernel Incremental revision model The market responses are the absolute two-year Treasury yield change and the absolute S&P 500 return in the USMPD statement window. Source percentage-point and percentage units are both divided by 100. We do not estimate a full-sample principal component, label the responses exogenous shocks, or infer direction from their absolute magnitudes. Controls for the release, document quantity and immediate response are retained when the revision term is added. rᵢ,ₜ = log(Pᵢ,ₜ / Pᵢ,ₜ₋₁) yᵢ,ₜ,ₕ = (1/h) Σⱼ₌₁ʰ r²ᵢ,ₜ₊ⱼ P denotes the adjusted close. The five-day horizon is primary; one and twenty days are secondary. Forecasts concern the average squared return, so all horizons remain on a daily second-moment scale. The h-day outcome uses only returns strictly after origin t. Adjusted closes reflect the downloaded historical adjustment series; they are not independently reconstructed historical data vintages.

Drag to resize

Response not available

Drag to resize