README.md

---
language:
- en
license: cc-by-4.0
tags:
- finance
- earnings-calls
- nlp
- market-microstructure
- sentiment-analysis
- benchmark
- finbert
- nbbo
pretty_name: 'MERIT: Microstructure and Earnings call Real-time Information Trajectories'
size_categories:
- 100K<n<1M
task_categories:
- text-classification
---

# MERIT: Microstructure and Earnings call Real-time Information Trajectories

## Overview

**MERIT** is a sentence-level benchmark that links earnings conference call language to sub-minute market microstructure responses. It repositions earnings call analysis from static, document-level summaries to **micro-temporal language trajectories** anchored to live market quote data.

Each observation pairs a unit of earnings call language—a presentation sentence or a Q&A exchange—with tick-level **NBBO (National Best Bid and Offer)** statistics measured in fixed windows immediately before and after that language unit is spoken.

The main benchmark covers **1,118 earnings call events** (regular-session and after-hours only) from **167 S&P 500 companies** spanning **2021–2023**. An additional 747 pre-market events are excluded from the main benchmark due to differences in pre-open quote liquidity, and are stored in `robustness/` for reference.

---

## Motivation

Prior earnings-call datasets focus on summarization, argumentation mining, or subjectivity annotation, but provide no market grounding. Finance datasets that measure market reactions operate at daily or longer horizons. MERIT fills this gap with three design principles:

1. **Sentence-level anchoring** — each utterance is temporally aligned to short-horizon quote responses, rather than aggregated at call or session level.
2. **Four representation regimes** — document-level aggregate, instantaneous utterance, cumulative level, and cumulative trajectory — allowing gains from each source of signal to be separately identified.
3. **Temporal generalisation** — a chronological train/test split (train: 2021–2022; test: 2023) prevents look-ahead bias and tests transfer to future market periods.

---

## Dataset Statistics

| Split | EC Events | Companies | Regular-session ECs | After-hours ECs |
|-------|-----------|-----------|--------------------|-|
| Train (2021–2022) | 729 | 159 | 382 | 347 |
| Test (2023) | 389 | 151 | 192 | 197 |
| Pre-market (excluded) | 747 | — | — | — |
| **Total (main benchmark)** | **1,118** | **167** | | |

Companies shared across train and test splits: **143**

### Anchor counts (Presentation)

| | Train | Test |
|--|-------|------|
| Raw sentence anchors | 88,650 | 50,891 |
| Selected anchors, post=30s | 24,101 | 13,896 |
| Selected anchors, post=300s | 5,346 | 2,968 |

### Anchor counts (Q&A)

| | Train | Test |
|--|-------|------|
| Raw QA-pair anchors | 9,699 | 5,600 |
| Selected anchors, post=30s | 8,600 | 5,050 |
| Selected anchors, post=300s | 3,895 | 2,311 |

---

## Company Selection

MERIT is anchored to S&P 500 constituents. Three filtering steps were applied:

1. **Sector exclusion** — Financials, Utilities, and Real Estate (GICS) are excluded due to systematic differences in disclosure conventions and trading environment.
2. **Quote coverage** — only firm-quarter events with a matched NBBO file in the NYSE TAQ database are retained.
3. **Structural validity** — only firms where transcript structure and quote coverage jointly support sentence-level alignment are included.

The retained firms span **eight GICS sectors**: Technology (44), Industrials (28), Healthcare (19), Consumer Cyclical (32), Consumer Defensive (14), Energy (16), Communication Services (7), Basic Materials (7).

---

## Language Features

### Presentation segment — FinBERT-tone

Each Presentation sentence is scored with **FinBERT-tone**, yielding:

- `finberttone_expected_value`: scalar tone = P(Positive) − P(Negative) ∈ [−1, 1]
- `finberttone_cumulative_tone`: running sum of expected value up to this sentence
- `finberttone_change_point`: 1 if this sentence starts a new sentiment regime (PELT + AIC penalty = 2); 0 otherwise

Median Presentation length: **164 sentences per call** (max: 626).

### Q&A segment — SubjECTive-QA

Each analyst–executive exchange is scored along **six dimensions** using SubjECTive-QA (Pardawala et al., NeurIPS 2024):
`assertive`, `cautious`, `optimistic`, `specific`, `clear`, `relevant`

Each dimension has three columns: `{dim}_negative_score`, `{dim}_neutral_score`, `{dim}_positive_score`.

Anchor timestamp: `Q_Timestamp` (question start — earliest moment the market can react).

Median Q&A depth: **14 pairs per call** (max: 65). Of the 1,118 main benchmark events, 1,032 contain at least one Q&A pair with valid quote coverage.

---

## Market Microstructure Targets

Three targets capture complementary dimensions of short-horizon quote dynamics, each expressed as the **change from the pre-window mean to the post-window mean**:

| Target | Description |
|--------|-------------|
| `∆BAS` | Change in relative bid–ask spread — trading frictions and liquidity costs |
| `∆QRF` | Change in quote revision frequency (NBBO updates/min) — price discovery intensity |
| `∆QVol` | Change in realised quote volatility (std of mid-price changes) — local price uncertainty |

Targets are binarised as extreme-event indicators:

```
Y = 1  if |∆metric| > τ₇₅   (75th percentile of training-split delta distribution)
Y = 0  otherwise
```

This focuses evaluation on the upper tail of microstructure changes and avoids label noise from near-zero fluctuations. **Balanced accuracy** is the primary metric (chance level = 0.500).

Majority-class baselines at post=300s: ∆BAS 82.6%, ∆QRF 65.3%, ∆QVol 77.7%.

---

## Window Design

For each anchor at time *t*, quote statistics are aggregated over:

```
|<── pre-window (fixed 30s) ──|── post-window (W) ──>|
t − 30s                       t                    t + W
```

Post-window widths **W ∈ {30, 60, 120, 300} seconds** → four files per anchor type.

### Non-overlapping anchor protocol

To prevent mechanical autocorrelation, anchors are selected greedily so that no two retained anchors share any portion of their post-window. For horizon W, each retained anchor must start at least W seconds after the previously retained anchor.

---

## Dataset Structure

```
data/
├── ec_calendar_sp500_with_sector_and_industry.csv    EC event calendar
│                                                     (1,118 main benchmark events;
│                                                      747 pre-market excluded)
├── benchmark/                        Full benchmark (all main-session ECs, pre split)
│   ├── benchmark_pre_{30,60,120,300}s.csv
│   ├── benchmark_qa_{30,60,120,300}s.csv
│   ├── benchmark_labeled/            Benchmark with microstructure labels
│   │   ├── benchmark_pre_{30,60,120,300}s.csv
│   │   └── benchmark_qa_{30,60,120,300}s.csv
│   ├── benchmark_coverage.csv        Per-EC anchor counts (valid / excluded)
│   └── excluded_anchors_{30,60,120,300}s.csv   Anchors excluded (n_ticks_post < 3)
└── benchmark_split/
    ├── train/                        Train set: 2021–2022
    │   ├── benchmark_pre_{W}s.csv
    │   └── benchmark_qa_{W}s.csv
    ├── test/                         Test set: 2023
    │   ├── benchmark_pre_{W}s.csv
    │   └── benchmark_qa_{W}s.csv
    └── robustness/                   Pre-market ECs (excluded from main benchmark)
        ├── pre_market_pre_{W}s.csv
        └── pre_market_qa_{W}s.csv
```

### File naming

```
benchmark_{anchor_type}_{post_window}s.csv
  anchor_type  :  pre  (presentation sentence)
               :  qa   (Q&A pair)
  post_window  :  30 | 60 | 120 | 300  (seconds)
```

---

## EC Calendar (`ec_calendar_sp500_with_sector_and_industry.csv`)

This file is the **master event index** for the benchmark. It is the single source of truth for EC start times, session labels, and company metadata.

| Column | Description |
|--------|-------------|
| `tic` | Ticker symbol |
| `name` | Company name |
| `sector` | GICS sector |
| `industry` | GICS industry |
| `year` | Earnings call year |
| `quarter` | Quarter (Q1–Q4) |
| `timestamp_start_utc` | EC start time in UTC (ISO 8601) |
| `timestamp_start_et` | EC start time in US/Eastern (ISO 8601) |

The `ec_session` column in all benchmark files is derived deterministically from `timestamp_start_et` in this calendar.

---

## Key Columns

### Identity columns (all files)

| Column | Description |
|--------|-------------|
| `tic` | Ticker symbol |
| `year` | Earnings call year |
| `quarter` | Quarter (Q1–Q4) |
| `anchor_type` | `pre` or `qa` |
| `anchor_id` | `section_id` (pre) or `QA_Index` (qa) |
| `post_window_sec` | Post-window width in seconds |
| `timestamp_anchor` | Seconds from EC start to this anchor |
| `ec_session` | `regular` / `after_hours` / `pre_market` / `non_trading` |

### Quote features (pre- and post-window)

Each appears with both `_pre` and `_post` suffix:

| Column | Description |
|--------|-------------|
| `bid_ask_spread_mean` | Mean relative bid–ask spread |
| `bid_ask_spread_std` | Std of bid–ask spread |
| `obi_mean` | Mean order book imbalance |
| `qrf_mean` | Mean quote revision frequency (ticks/min) |
| `quote_volatility_mean` | Mean quote volatility |
| `n_ticks` | Number of quote ticks in the window |

---

## Session Distribution

Session distribution computed over all main benchmark events (1,118) plus pre-market events (747):

| Session | ET Hours | Description |
|---------|----------|-------------|
| `regular` | 09:30–15:59 | Main trading session |
| `after_hours` | 16:00–19:59 | Post-close session |
| `pre_market` | 04:00–09:29 | Pre-open session (excluded from main benchmark) |

**Main benchmark** (`train/`, `test/`): `regular` + `after_hours` sessions only.

**Excluded from main benchmark**: `pre_market` sessions. NBBO data is systematically sparse before market open, making post-window targets less reliable as primary evaluation targets. Pre-market events are stored in `robustness/` for reference.

---

## Filtering

Anchors with **fewer than 3 quote ticks** in the post-window are excluded from the main benchmark. Excluded anchors are recorded in `excluded_anchors_{W}s.csv` for transparency.

Tick-level quality filters applied upstream:
- Crossed-NBBO ticks (bid > ask) discarded (~1.2% of ticks)
- Extreme-spread ticks (relative spread > 0.50) discarded (~1.3% of ticks)

---

## Baseline Results (Summary)

| Regime | Presentation BAcc (W=30s) | Q&A BAcc (W=30s) |
|--------|--------------------------|------------------|
| Document-level aggregate | 0.500 | — |
| Instantaneous utterance (Inst.) | 0.502 | 0.523 |
| Cumulative level (CL) | 0.508 | 0.552 |
| Cumulative trajectory (CT) | 0.520 | 0.560 |

Chance level = 0.500. Full results across all targets (∆BAS, ∆QRF, ∆QVol) and horizons (30s–300s) are reported in the accompanying paper.

---

## Data Release Scope

| Component | Released | Notes |
|-----------|----------|-------|
| Sentiment panel (FinBERT-tone + SubjECTive-QA scores) | ✅ | This repository |
| Benchmark split (pre-aggregated features) | ✅ | This repository |
| Raw earnings call transcripts | ❌ | Commercial licensing restrictions prohibit redistribution |
| Raw NYSE TAQ tick streams | ❌ | Commercial licensing restrictions prohibit redistribution |

---

## Reproducibility

The benchmark is **deterministically reproducible** given:
- The same EC calendar (start times in ET)
- The same tick-level quote panel
- `--min_ticks_post 3` (default filtering threshold)

Session labels depend only on EC start time (ET timezone).

---

## License

Released under [Creative Commons Attribution 4.0 (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/).
Pipeline code and baselines are released separately under the MIT License.

---

## Citation

> The accompanying paper is currently under review. Citation information will be added upon publication.

---

## References

- Pardawala et al. (2024). SubjECTive-QA: Measuring subjectivity in earnings call transcripts' QA through six-dimensional feature analysis. *Advances in Neural Information Processing Systems*, 37.
- Huang, Wang, & Yang (2023). FinBERT: A large language model for extracting information from financial text. *Contemporary Accounting Research*, 40(2), 806–841.
- Beaver (1968). The information content of annual earnings announcements. *Journal of Accounting Research*, 6, 67–92.
- Lee, Mucklow, & Ready (1993). Spreads, depths, and the impact of earnings information: An intraday analysis. *Review of Financial Studies*, 6(2), 345–374.
- Killick, Fearnhead, & Eckley (2012). Optimal detection of changepoints with a linear computational cost. *Journal of the American Statistical Association*, 107(500), 1590–1598.