Manuscript Skeleton

Working title:

Can Machine Learning Improve Earnings Event-Variance Trading? Evidence from U.S. Equity Options

This page is organized like a paper draft rather than a project plan. It keeps the current evidence conservative: the results are based on a no_nbbo_trade_proxy route and are not paper-grade executable trading results.

Abstract

This paper asks whether machine-learning models can improve trading decisions around option-implied earnings event variance mispricing. The object is not generic implied-volatility forecasting. Models forecast realized earnings-event variance, the market benchmark is option-implied event variance IVAR_event, and the tradable question is whether predicted mispricing improves premium-space trade selection after proxy transaction costs.

The current study uses a SEC-first earnings calendar and Massive market-data proxy route for U.S. single-name equity options from 2022-12-01 through 2025-12-31. The sample contains 810 BMO/AMC earnings events, of which 693 have a trade-proxy IVAR_event. The primary scientific target is close-to-open earnings jump variance (jump_c2o); the V1 proxy-PnL headline is close-to-close event variance (day_c2c); post-open digestion (reaction_o2c) is diagnostic.

In the current 2026-05-12 canonical tuned proxy package, the default fe_v2_sec_xbrl feature schema is not the sell: its strongest jump_c2o AUC is the Goyal-Saretto-style spread at 0.602, and the positive day_c2c ridge-flat sequence proxy PnL of about 19,918 USD remains diagnostic because the sequence gate does not pass. The same-code fe_v1_legacy ablation is stronger: LightGBM reaches jump_c2o AUC 0.677, XGBoost has best jump_c2o OOS R2 versus IVAR at 0.375, and LightGBM leads the day_c2c headline proxy strategy at about 53,664 USD net PnL. FT-Transformer refers to the validation-tuned tabular transformer specification; it trains but is not competitive. reaction_o2c is modeled as a diagnostic target, but its post-open realized variance is compared to full-event IVAR_event, so it is not a calibrated O2C mispricing or headline strategy result. The legacy in-repo proxy-Mamba rows are retired because they used a gated recurrent encoder rather than official mamba-ssm. The defensible conclusion is that a parsimonious tabular feature set shows preliminary cross-sectional ranking signal for earnings event-variance mispricing in a no-NBBO proxy sample, while FE V2 is currently a negative diagnostic result. Paper-grade claims require historical quote/NBBO or equivalent data, quote-based IVAR, and leg-level execution with realistic bid/ask crossing.

1. Introduction

Earnings announcements create scheduled jumps in uncertainty. Option prices embed a market forecast of this event variance, but the central empirical question is whether observable pre-event state and option-surface information can improve the cross-sectional ranking of event variance mispricing.

The paper asks:

Can models improve trading decisions around option-implied earnings event variance mispricing?

The realized-variance target system is:

RVAR_event_jump_c2o     = log(open_after / close_before)^2
RVAR_event_day_c2c      = log(close_after / close_before)^2
RVAR_event_reaction_o2c = log(close_after / open_after)^2

The market baseline is:

IVAR_event

The V1 tradable mispricing label is:

RVAR_event_day_c2c - IVAR_event

The trade rule is evaluated in premium space:

expected_strategy_edge_usd
  = expected_strategy_value_usd - market_entry_cost_usd

Forecast error is therefore only supporting evidence. The paper-facing result is whether a model improves ranking, edge selection, and proxy net performance in the tradable tail.

Contribution

The contribution is not a model-family claim. The intended contribution is narrower:

State and event-history features contain preliminary cross-sectional signal for earnings event-variance mispricing beyond market-implied IVAR and simple historical baselines.

The model comparison is outcome-dependent. If the sequence suite passes the diagnostic gate, ordered pre-event proxy-surface paths may contain incremental information. If LightGBM/XGBoost win, event-level nonlinear tabular interactions are sufficient for the current proxy data. If IVAR wins after costs, the evidence supports a hard-to-beat earnings option market. The current same-code ablation favors a parsimonious FE V1 tabular interpretation, not an FE V2 or deep-sequence headline.

Related Literature and Positioning

Literature stream	Closest role in this paper
Earnings option pricing and scheduled jumps	Motivates separating event variance from total short-dated variance.
Earnings straddle-return studies	Motivates testing whether predicted event variance mispricing maps into option strategy returns.
RV-IV spread and option-return predictability	Provides required classical benchmarks, including Goyal-Saretto-style spread signals.
Empirical asset pricing with ML	Sets the discipline: out-of-sample ranking and economic value matter more than in-sample fit.
Surface and sequence models	Motivates FT-Transformer and official mamba-ssm diagnostics, but only after strong tabular baselines.

The paper differs from average-return earnings straddle studies by asking whether models sort events by expected event variance mispricing and whether that sorting survives proxy costs.

2. Data

2.1 Sources and Execution Grade

The current data route uses official SEC filings for event identification and Massive market-data proxies for prices:

• SEC EDGAR 8-K / 8-K/A Item 2.02 filings and SEC primary-document text validation.
• SEC company ticker metadata for eligible common-equity-like single-name underlyings.
• Massive options day aggregates for universe liquidity ranking, contract discovery, close-trade-implied IV proxies, and daily sequence features.
• Massive option contract reference metadata for multiplier and deliverable validation.
• Massive underlying day aggregates for event returns and vendor OHLC opens.
• Massive option one-second aggregates for entry prices, C2C exit marks, and C2O/O2C post-open diagnostic marks.
• FRED VIXCLS for prior-close daily market-state controls.

All current option second aggregates are trade OHLCV bars. They are not quote midpoints, bid/ask records, OPRA, or NBBO. The current panel grade is no_nbbo_trade_proxy; paper_grade=false.

2.2 Sample and Universe

The active proxy run covers 2022-12-01 through 2025-12-31. The target paper range remains 2013-2025, but that requires upgraded historical option data or a separate licensed route.

The single-name universe is dynamic. Each month, the pipeline ranks eligible underlyings by trailing six-month option premium dollar volume:

option_premium_dollar_volume = option_price * contract_volume * 100

ETF, fund, trust, ETN, index, volatility, commodity, and other non-single-name symbols are excluded before the top-50 ranking. BMO and AMC events are retained; DMH and unknown timing are excluded from the main sample.

2.3 Current Data Coverage

Measure	Value
Dynamic-calendar rows	1,054
BMO/AMC main-sample candidates	810
Trade-proxy event-panel rows	810
Events with C2C rvar_event alias	801
Events with trade-proxy IVAR_event	693
Proxy contract candidates	12,038
Contracts with usable pre-cutoff proxy price	10,165
Contracts with no trade in cutoff window	1,873
Contracts with local IV proxy	10,138
Main DTE 5-14 contracts	5,098
Robustness DTE 3-21 contracts	12,038
Proxy straddle diagnostic rows	779

IVAR failure diagnostics:

Failure reason	Events
No two event-covering expiries	103
Nonmonotone total variance	7
Negative extracted IVAR	7

The event panel is large enough for proxy-stage model comparison, but IVAR coverage is still a material screen: 117 of 810 events lack a usable trade-proxy IVAR.

3. Methods

3.1 Pipeline

The pipeline separates event discovery, market-data construction, feature engineering, model training, and proxy backtesting. The diagram keeps the execution caveat explicit: current prices are trade-aggregate proxies.

flowchart TB
  sec["SEC EDGAR 8-K Item 2.02\nprimary filing text"] --> cal["BMO/AMC event calendar"]
  secmeta["SEC company ticker metadata"] --> universe["Eligible monthly top-50\nsingle-name option universe"]
  optday["Massive option day aggregates"] --> universe
  universe --> cal
  cal --> align["Event alignment\nBMO/AMC entry timestamps"]
  stockday["Underlying day aggregates"] --> targets["C2O/C2C/O2C targets"]
  align --> targets
  optref["Option reference metadata"] --> contracts["Standard contract validation"]
  optday --> contracts
  contracts --> ivar["IVAR_event\n2-expiry event variance extraction"]
  optsec["Option one-second trade aggregates\nnot quote, not NBBO"] --> entry["Pre-cutoff entry VWAP"]
  optsec --> exit["C2C exit preclose VWAP\nC2O/O2C post-open VWAP diagnostics"]
  ivar --> features["Feature matrix and sequences"]
  targets --> features
  entry --> features
  features --> models["Benchmarks, tabular models,\nFT-Transformer, sequence diagnostics"]
  models --> metrics["Forecast, ranking,\nproxy strategy metrics"]
  exit --> metrics
  metrics --> docs["Tables, figures,\nproxy report, docs"]

3.2 Event Alignment and Leakage Control

Feature construction uses a hard as-of gate:

feature_asof_timestamp <= event_entry_timestamp

AMC events enter before the announcement-date close. BMO events enter before the previous trading-day close. Vendor daily OHLC opens are used for C2O target construction and labeled as vendor regular OHLC assumptions, not verified auction prints.

3.3 IVAR Construction

For two event-covering expiries, total ATM implied variance is:

w(T) = sigma_ATM(T)^2 * T

The implied event variance is extracted as:

IVAR_event = (T2*w1 - T1*w2) / (T2 - T1)

Negative extracted event variance and nonmonotone total variance are excluded from tradable samples and reported as diagnostics.

3.4 Features

The default research feature schema is fe_v2_sec_xbrl; fe_v1_legacy is kept only for same-code ablations. The resolved run-level allowlist is artifacts/modeling/feature_schema_report.csv, and only model_feature=true rows enter trainable models. FE V2 removes raw numeric identifiers, raw year/month, exit/outcome/PnL fields, and post-event labels from the model matrix. The signal timestamp is event_entry_timestamp, so completed pre-cutoff entry-window features are valid under the current protocol.

The feature matrix combines event-level state, realized history, option-surface proxies, market controls, and sequence inputs:

• IVAR_event, ATM IV, term spread, skew, butterfly/concavity proxies.
• Option activity and liquidity measures.
• RV5/RV20/RV60, last-four earnings history, and strict point-in-time same-ticker rolling earnings-history distributions.
• BMO/AMC timing and universe rank.
• Prior-close VIX level, changes, percentile, and regime.
• SPY/QQQ controls when available.
• SEC CompanyFacts XBRL fundamentals with conservative as-of gating: use acceptanceDateTime <= feature_asof_timestamp when mapped, otherwise allow only filed < feature_asof_date.
• Train-fitted cross-sectional z-score/rank transforms; locked-test distribution is never fit.
• Single-name 1/3/5/10-day run-up, weak delta-grid, and RND-like *_proxy features from trade-aggregate implied surfaces. These are not quote surfaces, NBBO surfaces, or paper-grade RND estimates.
• Daily 20-step close-trade-implied option-surface sequences.
• Hybrid 31-step sequences with 19 daily states and 12 entry-day five-minute trade-aggregate proxy bins.

Sequence coverage is 678 eligible events out of 810. The default drop rate is 16.3%, so sequence results are diagnostic in the current run.

3.5 Models

Family	Models	Purpose
Market benchmark	Market-implied IVAR	Central level and no-edge baseline.
Historical baselines	Last-four RVAR, last-four IVAR	Tests whether simple earnings history is enough.
Classical mispricing benchmark	Goyal-Saretto-style RV-IV spread	Required option-return predictability comparator.
Linear tabular	Elastic Net	Sparse linear event-level benchmark using sklearn ElasticNetCV.
Nonlinear tabular	LightGBM, XGBoost	Main current contenders with validation-only tuning.
Ensemble	LightGBM/XGBoost rank-average	Robustness ensemble built from tuned base forecasts.
Neural tabular	FT-Transformer	Validation-tuned deep tabular comparator.
Sequence diagnostics	Ridge-flat sequence aggregates, BiGRU 5-seed, official bidirectional mamba-ssm 5-seed, attention pooling, non-causal dilated CNN, mask-only and time-shuffle controls	Tests whether ordered pre-event paths add value.

The canonical protocol is the tuned-only proxy protocol. Tuning uses only train and locked-validation rows, selects on validation jump_c2o predicted-edge AUC with top-decile precision and RMSE tie-breakers, then refits on train+validation before a single locked-test evaluation. Paired original tabular rows and single-seed BiGRU/Mamba rows are intentionally excluded from the current artifacts.

The full sequence diagnostic suite is diagnostic-grade in the current sample. It runs jump_c2o, day_c2c, and reaction_o2c for ridge-flat, BiGRU 5-seed, official mamba-ssm 5-seed, attention pooling, non-causal dilated CNN, mask-only, and time-shuffle controls. The official Mamba wrapper is bidirectional over completed pre-entry tokens and is therefore a non-causal encoder of the pre-event path, not a post-entry leakage channel.

3.6 Splits, Strategy, and Metrics

The current proxy run uses chronological event-level 70/15/15 train, validation, and test splits. The split unit is event_id, so C2O/C2C/O2C rows for the same event cannot cross splits.

The V1 strategy headline is day_c2c only. Entry uses per-leg option VWAP over the final 900 seconds before cutoff. The primary C2C exit uses same-contract option VWAP over the final 15 minutes before the exit-date close. C2O and O2C option-PnL rows are diagnostic decompositions.

Performance metrics:

Metric family	Metrics
Forecast	MAE, RMSE, QLIKE diagnostic, OOS R2 versus IVAR
Ranking and mispricing	AUC, Brier, calibration, top-decile precision, edge-decile monotonicity
Strategy	Gross/net proxy PnL, return on premium/capital, Sharpe, Sortino, max drawdown, hit rate, average win/loss, cost sensitivity
Risk and coverage	IVAR failure counts, sequence drop rate, high sequence-selection risk, extreme prediction diagnostics

4. Results

paper_plan.md is the manuscript skeleton, not the full result ledger. The complete C2C/C2O/O2C tables, figures, diagnostics, and interpretation now live in Results Snapshot. This section records the intended paper-facing organization and the selected headline excerpt.

4.1 Feature-Schema Ablation Headline

The active default artifacts use fe_v2_sec_xbrl, but the same-code ablation is negative for FE V2. Forecast and ranking columns below use jump_c2o unless the target column states otherwise. Strategy columns use day_c2c only for the headline proxy-PnL interpretation; C2O/O2C premium-space rows are diagnostic.

Feature schema	Target	Best AUC model	Best AUC	Best OOS R2 model	Best OOS R2 vs IVAR	Best headline/diagnostic PnL model	Best net PnL
fe_v1_legacy	jump_c2o	LightGBM	0.677	XGBoost	0.375	Official mamba-ssm 5-seed, C2O intrinsic diagnostic	28,898
fe_v1_legacy	day_c2c	LightGBM	0.925	XGBoost	0.574	LightGBM, C2C headline	53,664
fe_v1_legacy	reaction_o2c	Ridge-flat sequence	0.799	XGBoost	0.949	FT-Transformer, O2C diagnostic	643
fe_v2_sec_xbrl	jump_c2o	Goyal-Saretto spread	0.602	LightGBM	0.203	Official mamba-ssm 5-seed, C2O intrinsic diagnostic	28,898
fe_v2_sec_xbrl	day_c2c	Ridge-flat sequence	0.636	Ridge-flat sequence	0.264	Ridge-flat sequence, C2C headline	19,918
fe_v2_sec_xbrl	reaction_o2c	Ridge-flat sequence	0.799	LightGBM/XGBoost ensemble	0.945	FT-Transformer, O2C diagnostic	753

The central result is not that the richest feature schema wins. The current paper-facing signal screen is the parsimonious FE V1 tabular result; FE V2, FT-Transformer, and sequence rows remain diagnostic until follow-up feature audits show otherwise.

4.2 Results Order for the Manuscript

The results section should be written in this order:

1. C2C: forecast/ranking table, premium-space proxy strategy table, edge-decile strategy figure, and interpretation. This is the only V1 proxy-PnL headline.
2. C2O: forecast/ranking table, forecast/ranking/calibration/monotonicity figures, post-open diagnostic strategy table, and interpretation. These rows support the scientific ranking story but are not headline execution claims.
3. O2C: forecast/ranking table, O2C forecast/ranking/strategy/scale figures, post-open premium-space diagnostic strategy table, scale-mismatch diagnostic, and interpretation. O2C is diagnostic because IVAR_event is a weak full-event comparator for post-open realized variance.
4. Sequence diagnostics: ridge-flat, BiGRU 5-seed, official bidirectional mamba-ssm 5-seed, attention pooling, non-causal dilated CNN, mask-only, and time-shuffle controls. The current conclusion is negative: ordered proxy-surface paths do not beat tabular aggregates or the controls.
5. Robustness and inference: cost sensitivity, clustered forecast-loss inference, calibration, and QLIKE caveats. These support a conservative tabular-model claim and do not replace quote/NBBO execution evidence.

5. Limitations

The current evidence is not final paper-grade executable evidence. The main limitations are:

Limitation	Consequence
No historical bid/ask, quote midpoint, OPRA, or NBBO records	Cannot claim full-spread executable strategy performance.
Option second aggregates are trade OHLCV bars	IV surfaces and strategy marks are trade-price proxies.
Current sample starts in 2022	Does not yet cover the target 2013-2025 paper window.
117 of 810 events lack usable trade-proxy IVAR	IVAR coverage is a material sample screen.
Sequence eligibility is 678 of 810 events	Sequence results carry selection risk and are diagnostic.
C2O/O2C option PnL is diagnostic	V1 tradable mispricing headline remains day_c2c.
Proxy haircut cost model	Full bid/ask crossing remains future paper-grade work.

Paper-grade claims require historical quote/NBBO or equivalent data, quote-based IVAR, leg-level execution with realistic bid/ask crossing, DTE and liquidity robustness, and clustered or bootstrap inference.

6. Conclusion

The current proxy-stage evidence supports a disciplined, limited conclusion. The same-code ablation says a parsimonious FE V1 tabular feature set improves the ranking of earnings event-variance mispricing relative to market IVAR and simple historical benchmarks, and maps this ranking signal into positive day_c2c premium-space proxy economics. The richer FE V2 default is currently a negative diagnostic result rather than a headline improvement.

The result is not a final execution claim. It is a credible signal-screening result that justifies either a paper-grade quote/NBBO extension or a conservative proxy-stage manuscript. If the next paper-grade route confirms the same ranking and cost robustness under bid/ask execution, the sell is an earnings event-variance mispricing paper. If it does not, the paper still has a useful negative result: market-implied event variance and strong tabular baselines are hard to beat under realistic earnings-option frictions.

Appendix Plan

Appendix	Contents
A. Literature and Positioning	Earnings option pricing, event volatility, option-return predictability, and ML comparisons.
B. Universe Construction	Monthly top-50 membership, turnover, exclusions, and liquidity distributions.
C. Event Calendar Audit	SEC accessions, timing flags, text validation, and BMO/AMC exclusions.
D. IVAR Diagnostics	Expiry selection, DTEs, total variances, negative IVAR, and nonmonotone failures.
E. Feature Schema	Event-level, VIX, SPY/QQQ, daily sequence, hybrid sequence, and as-of timestamps.
F. Model Configuration	Splits, hyperparameters, seeds, training status, and fit diagnostics.
G. Robustness and Inference	DTE windows, liquidity buckets, timing splits, ticker/year concentration, clustered SEs, and bootstrap checks.