Monetary Policy Expectations (MPE) Index

• The MPE Index is a text-derived measure that quantifies forward-looking monetary policy expectations by processing FOMC statements and social media narratives.
• One construction uses regression-based projections of market surprises onto differences in text-implied expectations, while the other aggregates weighted investor sentiment from StockTwits.
• Both variants isolate the informational component of policy moves, offering insights into market reactions and predictive power for asset yields and crypto returns.

Searching arXiv for the cited papers to verify metadata and ensure current grounding. The Monetary Policy Expectations (MPE) Index denotes a text-derived measure of monetary-policy expectations, but the term is used for two distinct constructions in the recent literature. In one formulation, derived from Federal Open Market Committee (FOMC) statements and contemporaneous New York Times coverage, the index is the fitted informational-content component of a high-frequency monetary policy shock, identified as the projection of market movements onto the difference between FOMC and private-sector expectations (Cai et al., 2021). In another formulation, developed for crypto-financial applications, the index is a weekly, engagement-weighted average of LLM-classified hawkish and dovish narratives extracted from StockTwits messages tagged with $FED` and `$MACRO, designed to capture ex-ante market expectations rather than realized policy implementation (Nicolas et al., 9 Apr 2026). Across both uses, the core object is forward-looking monetary-policy information embedded in text, but the sampling frequency, unit of measurement, data source, and econometric role differ materially.

1. Conceptual scope and definitional variants

The term MPE Index is not attached to a single canonical formula across the two papers. In the FOMC-announcement framework, the paper itself refers to the “information/news component” rather than a named index, and the index can be defined from the method as

$MPE_t^{R} := \hat{\theta} D_t,$

where $D_t$ is the difference between text-implied FOMC expectations and text-implied private-sector expectations, and $\hat{\theta} D_t$ is the fitted component of a high-frequency shock variable such as FF4 or PNS explained by that difference (Cai et al., 2021). In the StockTwits-based framework, the MPE index is explicitly a high-frequency, text-derived measure of ex-ante market expectations about monetary policy, aggregated at weekly resolution from LLM-classified investor communications (Nicolas et al., 9 Apr 2026).

These two constructions share a common interpretive axis: both attempt to isolate what market participants expect monetary policy to mean, rather than merely recording realized changes in the policy rate. The overlap is conceptual rather than operational. One series is event-based and denominated in basis points (bps) of a shock variable explained by information disclosure; the other is a weekly weighted average of ordinal hawkish/dovish scores. A plausible implication is that “MPE Index” is best treated as a family of text-based expectation measures rather than a single standardized macro-financial indicator.

Variant	Core definition	Frequency / units
FOMC informational-content measure	Fitted value from projecting $\Delta R_t$ onto $D_t$	Scheduled FOMC meetings / bps
StockTwits narrative measure	Engagement-weighted average of LLM-assigned scores $s \in \{-2,-1,0,+1,+2\}$	Weekly / index level

The distinction matters substantively. The FOMC-based measure is an identification device for decomposing high-frequency policy surprises into information and monetary components. The StockTwits-based measure is a narrative indicator used as an explanatory and predictive variable in VAR, VMD, and LSTM-SHAP analyses.

2. Event-based MPE from FOMC and private-sector text

In the 2021 framework, the informational content of FOMC announcements is identified by modeling the expectations of the FOMC and private-sector agents using computational linguistic tools on FOMC statements and New York Times articles, then projecting high-frequency financial-market movements onto differences in those expectations (Cai et al., 2021). The sample covers January 2000 to March 2014, includes 106 scheduled meetings, and excludes unscheduled meetings.

The text pipeline uses BERT (bert-base-uncased; Hugging Face implementation) to generate document embeddings. Preprocessing removes numbers, dates, and stop-words (gensim list), tokenizes with the Hugging Face default, splits documents into 256-token windows with 10-token overlap, and computes the document embedding as the mean of the BERT [CLS] token across windows. The output dimension is 768. The description explicitly states that no topic models or sentiment scorers are used; the framework relies on contextual embeddings and linear mapping.

These embeddings are mapped to expected FFR decisions through two linear elastic-net models:

• $f^{FOMC}(X_t^{FOMC}) \approx E_t^{FOMC}(FFR)$
• $f^{NYT}(X_t^{NYT}) \approx E_t^{PS}(FFR)$

The target variable is the FOMC’s target policy rate decision $FFR_t$ at meeting $MPE_t^{R} := \hat{\theta} D_t,$0 in bps. The elastic net is selected for transparency and regularization, with tuning parameters $MPE_t^{R} := \hat{\theta} D_t,$1 and $MPE_t^{R} := \hat{\theta} D_t,$2, and the models are chosen to maximize the $MPE_t^{R} := \hat{\theta} D_t,$3 in Stage 3, emphasizing predictive ability and interpretability over black-box complexity. The framework assumes the conditional mean restrictions

$MPE_t^{R} := \hat{\theta} D_t,$4

and linearity in embeddings for tractability.

The expectations difference is

$MPE_t^{R} := \hat{\theta} D_t,$5

This quantity is the driver of the identification strategy. It is not itself yet the index; rather, it is the text-based gap between what the FOMC statement implies and what private-sector reporting implied before or at the meeting.

3. Formal identification and decomposition in the FOMC framework

The high-frequency shock variable is denoted $MPE_t^{R} := \hat{\theta} D_t,$6 and may be a Fed Funds Futures surprise (FF4) or the Policy News Shocks (PNS) composite. Identification proceeds via the Stage 3 regression

$MPE_t^{R} := \hat{\theta} D_t,$7

The informational-content component is the fitted value from projecting $MPE_t^{R} := \hat{\theta} D_t,$8 onto $MPE_t^{R} := \hat{\theta} D_t,$9, written in vector form as

$D_t$0

In the scalar-$D_t$1 case, this is equivalent to $D_t$2 up to an intercept (Cai et al., 2021).

On that basis, the MPE Index for shock $D_t$3 at meeting $D_t$4 is defined as

$D_t$5

The residual

$D_t$6

is the policy action residual, yielding the decomposition

• News$D_t$7 or information component: $D_t$8
• Monetary$D_t$9 or action component: $\hat{\theta} D_t$0

The sign convention is explicit. A positive MPE indicates that the FOMC is more hawkish than the private sector, so markets learn “tightening” or stronger inflation outlook from the announcement, moving short rates up. A negative MPE indicates dovish information and lower short rates. No additional controls are required in Stage 3 because the theory posits that the information effect is proportional to $\hat{\theta} D_t$1; the description further states that omitted-variable bias arises if one regresses on $\hat{\theta} D_t$2 alone without $\hat{\theta} D_t$3.

The high-frequency identification details are also tightly specified. PNS is computed over 30-minute windows around the statement release and is the first principal component of five unanticipated rate changes involving current and near-term Fed funds expectations and 2–4 quarter eurodollar rates. FF4 follows the published Gertler–Karadi construction adopted by the paper. Validation uses daily changes in nominal Treasury yields (3m, 1y, 2y, 5y, 10y, 20y) and TIPS real yields (2y, 5y, 10y, 20y). Following Nakamura–Steinsson, some displays omit July 2008 to July 2009 to avoid crisis-period confounds.

4. Weekly LLM-based MPE from market messages

The 2026 construction defines the MPE index as a weekly, engagement-weighted measure of hawkish and dovish monetary-policy expectations extracted from 118,479 StockTwits messages over 2014-09 to 2025-02 (Nicolas et al., 9 Apr 2026). The corpus consists solely of messages tagged $\hat{\theta} D_t$4MACRO, with no addition of news wires, central-bank speeches, analyst notes, or multilingual sources.

Classification is performed using Mistral-7B (Jiang et al., 2023), run via the ollama Python stack. Each message is assigned one of five ordered categories:

• Very Hawkish $\hat{\theta} D_t$5
• Hawkish $\hat{\theta} D_t$6
• Neutral $\hat{\theta} D_t$7
• Dovish $\hat{\theta} D_t$8
• Very Dovish $\hat{\theta} D_t$9

The sign convention is therefore the reverse of the event-based bps interpretation: here, hawkishness is negative and dovishness is positive. Ambiguous or descriptive messages default to Neutral (0), intensity is encoded through the five-point scale, and no continuous $\Delta R_t$0 scoring is used. The paper does not report a human-labeled evaluation set, accuracy/precision/recall, confusion matrices, or inter-annotator agreement.

Let $\Delta R_t$1 denote the LLM-assigned score for message $\Delta R_t$2, and let $\Delta R_t$3 denote engagement-based weight derived from likes and reshares. For week $\Delta R_t$4, with message set $\Delta R_t$5, the index is

$\Delta R_t$6

The series is aligned to week-ending Friday. No additional normalization, such as z-scoring, and no smoothing, such as exponential moving averages, is applied.

The descriptive statistics reported for 2014–2025 are:

• mean −0.03
• standard deviation 0.09
• skewness −1.37
• kurtosis 13.56

The paper also notes that Neutral comprises 95.96% of messages, while extreme views attract more engagement, with Very Hawkish averaging 1.32 interactions against 0.31 for Neutral. Beyond engagement-weighting, there are no additional credibility filters or source-level reweighting. The paper does not report explicit deduplication or spam filtering for index construction.

Although the empirical analysis uses the aggregate index, the description provides a conceptual decomposition into hawkish and dovish contributions:

$\Delta R_t$7

5. Empirical behavior and econometric roles

In the FOMC-based framework, the central empirical object is the Stage 3 regression of $\Delta R_t$8 on $\Delta R_t$9. The reported pre-ZLB results are:

• PNS: $D_t$0; $D_t$1; $D_t$2
• FF4: $D_t$3; $D_t$4; $D_t$5
• FFR: $D_t$6; $D_t$7; $D_t$8

For the full sample (Jan 2000–Mar 2014), the reported result for FF4 is

• FF4: $D_t$9; $s \in \{-2,-1,0,+1,+2\}$0; $s \in \{-2,-1,0,+1,+2\}$1

The description states that PNS and FFR coefficients are positive but less significant in the full sample (Cai et al., 2021). The interpretation is direct: positive $s \in \{-2,-1,0,+1,+2\}$2 implies that when FOMC information points to a higher optimal policy rate than markets expected, FF4 or PNS rises. The recovered series is therefore treated as economically meaningful and statistically significant for FF4, and for PNS pre-ZLB.

The time-series behavior is illustrated with specific episodes. September 18, 2007 is described as a large negative MPE, reflecting markets inferring a deteriorating macro outlook from an aggressive cut. March 18, 2008 is a large positive MPE, associated with inflation concerns. January 30, 2002 is also a positive MPE, interpreted as communication of stabilization prospects.

The decomposition into Monetary$s \in \{-2,-1,0,+1,+2\}$3 and News$s \in \{-2,-1,0,+1,+2\}$4 is then used to study yield effects. For PNS, News$s \in \{-2,-1,0,+1,+2\}$5 significantly raises short-term nominal yields, including 3m: 1.208 bps per unit of News$s \in \{-2,-1,0,+1,+2\}$6; highly significant, while longer real rates show limited or negative responses, including TIPS 5–10y negative and significant. For FF4, News$s \in \{-2,-1,0,+1,+2\}$7 significantly affects 3m nominal yields, whereas Monetary$s \in \{-2,-1,0,+1,+2\}$8 explains most mid- to long-term nominal yield movements. The stated economic meaning is that the information effect primarily moves short maturities, consistent with revelation about near-term macro conditions and policy stance.

In the StockTwits-based framework, the MPE index is not a decomposition term but a predictive explanatory variable. It is designed to distinguish ex-ante expectations from ex-post policy moves, and the reported correlations are:

• $s \in \{-2,-1,0,+1,+2\}$9
• $f^{FOMC}(X_t^{FOMC}) \approx E_t^{FOMC}(FFR)$0
• $f^{FOMC}(X_t^{FOMC}) \approx E_t^{FOMC}(FFR)$1
• $f^{FOMC}(X_t^{FOMC}) \approx E_t^{FOMC}(FFR)$2
• $f^{FOMC}(X_t^{FOMC}) \approx E_t^{FOMC}(FFR)$3

The econometric specification is a bivariate VAR for Bitcoin returns with lag order chosen by AIC, testing the null that lagged MPE terms do not enter the return equation. The reported result is that the MPE index Granger-causes Bitcoin returns at lags 3–5, with p-values 0.069, 0.092, 0.089. In the frequency domain, Variational Mode Decomposition (VMD) splits variables into three intrinsic mode functions (IMFs): long-term (IMF1), medium-term (IMF2), and high-frequency (IMF3). The reported scale-specific result is that MPE IMF2 significantly predicts BTC IMF3 at lags 2 and 3, with $f^{FOMC}(X_t^{FOMC}) \approx E_t^{FOMC}(FFR)$4 and $f^{FOMC}(X_t^{FOMC}) \approx E_t^{FOMC}(FFR)$5 (Nicolas et al., 9 Apr 2026).

6. Modeling extensions, interpretation, and limitations

The 2026 study supplements linear causality tests with an LSTM that predicts one-week-ahead BTC return using 18 macro/market predictors plus lagged BTC, trained in an expanding-window walk-forward scheme with 4 folds, Optuna tuning, MSE loss, and RMSE and MAE evaluation metrics (Nicolas et al., 9 Apr 2026). Reported performance is:

• LSTM overall RMSE 0.0896 (MAE 0.0709)
• ARIMA overall RMSE 0.0873 (MAE 0.0666)

Diebold–Mariano tests show no significant difference in 1-step predictive accuracy across folds ($f^{FOMC}(X_t^{FOMC}) \approx E_t^{FOMC}(FFR)$6), and the LSTM is retained for its ability to capture non-linear, regime-dependent interactions. SHAP then ranks MPE as the third most important predictor on average, behind Google “inflation” and FFR, with global importances 0.008, 0.006, and 0.005, respectively. Temporal attribution assigns notable SHAP mass at $f^{FOMC}(X_t^{FOMC}) \approx E_t^{FOMC}(FFR)$7, $f^{FOMC}(X_t^{FOMC}) \approx E_t^{FOMC}(FFR)$8, $f^{FOMC}(X_t^{FOMC}) \approx E_t^{FOMC}(FFR)$9, $f^{NYT}(X_t^{NYT}) \approx E_t^{PS}(FFR)$0, described as “predictive memory.” Stratification by FFR regime (Rising/Flat/Falling) shows a robust negative slope between hawkish MPE readings and SHAP contributions to Bitcoin returns, including in the Flat regime, which the paper interprets as narrative effects independent of realized policy adjustments.

The FOMC-based measure is explicitly compared with the high-frequency shock literature. It is said to refine standard high-frequency shocks (FF4, PNS) by isolating the central bank information effect via text-based expectation differences, and to complement prior decompositions such as Jarociński–Karadi’s monetary vs information shocks, but by directly modeling private vs FOMC expectations from text rather than imposing sign/moment restrictions (Cai et al., 2021). The incremental explanatory power is summarized as nontrivial $f^{NYT}(X_t^{NYT}) \approx E_t^{PS}(FFR)$1 for FF4, reaching 8.5% pre-ZLB and 7.3% full sample.

Both constructions also carry clear caveats. In the FOMC case, the limitations include the rational expectations conditional on text embeddings assumption, linearity in Stage 2 and Stage 3, possible loss of information from removing numbers/dates, the use of general-corpus BERT, the small sample of 106 scheduled meetings, exclusion of press conference vs statement effects, exclusion of unscheduled meetings, and possible post-crisis underperformance at the zero lower bound. The description adds that the simple linear specification is interpreted as producing lower bounds.

In the StockTwits case, the stated limitations include classification risk due to the absence of a human-labeled validation set, coverage bias because StockTwits skews toward active retail or semiprofessional investors focused on U.S. policy, the dominance of Neutral messages, the possibility that engagement-weighting amplifies echo-chamber effects, sensitivity to weekly aggregation and Friday-close alignment, and the possibility that LLM updates or platform behavior changes may shift classification boundaries. External validity beyond U.S. monetary policy narratives and beyond the studied crypto context is not assessed.

A common misconception is that an MPE index simply restates realized policy-rate changes. Neither construction is designed that way. The FOMC-based measure isolates the fitted information disclosure component of a high-frequency surprise, while the StockTwits-based measure is explicitly distinguished from realized EFFR changes by construction and by weak correlation. Another potential misconception is that the two indices are directly interchangeable. They are not: one is an event-study projection in basis points, the other a weekly narrative score. What they share is the objective of quantifying expectations embedded in monetary-policy communication and market discourse rather than treating observed rate changes as sufficient statistics for policy transmission.