Guruprasath12/lie-detect

Earnings miss predictor (track record, sector, analysts, price, guidance)

Predicts the probability that a US company will miss the consensus EPS estimate at its next quarterly earnings announcement, about 90 days ahead, from:

• its own recent earnings surprises (5 numbers),
• its GICS sector (10 sectors; Real Estate not supported),
• analyst actions in the 90 days before: rating changes and price targets raised or cut (4 numbers),
• the stock's run-up vs the S&P 500 (SPY) over 1 and 3 months, and its volatility (3 numbers), and
• what its latest earnings press release says about guidance: raised, cut, maintained, narrowed, introduced or withdrawn, and how hedged the outlook is (10 numbers from regex rules).

It is a logistic regression (32 features → 1 logit) wrapped as a PyTorch module. It was built as a learning project; the GitHub repo has the full pipeline and the docs, including every idea that did not work.

What the evaluation says: most of the skill comes from the track record and the sector. On companies the model never saw, analyst, price and guidance features add +0.020 AUC over those two [+0.010, +0.030]. Forward in time the gain is +0.010 [-0.006, +0.027]: it helped in 2023 and 2024 but not in 2025–26 (table below).

Usage

# pip install torch huggingface_hub numpy pandas yfinance
# download miss_model.py, guidance_rules.py and market_features.py from this repo
from miss_model import HubMissModel

model = HubMissModel.from_pretrained("Guruprasath12/lie-detect")
p_miss = model.predict(
    past_surprises=[4.2, -1.5, 2.8, 6.0],   # EPS surprise % of earlier announcements, oldest → newest
    sector="Industrials",                     # one of model.sectors
    release_text=open("press_release.txt").read(),   # latest earnings press release (8-K Exhibit 99.1)
    ticker="CAT",                             # analyst actions and prices are fetched from Yahoo Finance
    announcement_date="2026-11-04",           # the announcement being predicted (YYYY-MM-DD)
)

Pass market_inputs=(analyst_actions, prices, spy_prices) instead of ticker to use your own data (format: market_features.market_vector). model.predict_proba(X) takes a raw feature matrix in model.features order; the model standardises inputs with training-set statistics stored as buffers.

History features (computed from announcements before the one predicted):

Feature	Definition
prev_miss	1 if the company missed the EPS estimate at its previous announcement, else 0
prev_surprise	previous announcement's EPS surprise in %, clipped to [-100, 100]
miss_rate_4q	share of misses over the last 4 announcements
mean_surprise_4q	mean EPS surprise % over the last 4 announcements
std_surprise_4q	std of EPS surprise % over the last 4 announcements (0 if < 2)

Sector: one-hot over model.sectors (Communication Services, Consumer Discretionary, Consumer Staples, Energy, Financials, Health Care, Industrials, Information Technology, Materials, Utilities).

Analyst and price features: market_features.market_vector. Windows end 2 days before the announcement date, so the release and the reactions to it never leak in. Analyst actions (last 90 days): log count, upgrades − downgrades, share of price targets raised − lowered, mean % target change. Price: 1- and 3-month return minus SPY's, 3-month volatility.

Guidance features: guidance_rules.guidance_vector(text) finds sentences mentioning guidance / outlook / forecast (minus boilerplate) and takes the direction verb closest to that noun. About 85% precise on spot checks. Use the text of the press release, without tables.

Training data

• 6,916 (press release, next announcement) pairs from 188 S&P 500 companies, 2017–2026; 16.8% misses. The release is filed on announcement day; it is paired with the next announcement (83–98 days later), never its own.
• Text: 8-K Item 2.02 Exhibit 99.1 from SEC EDGAR. Labels, analyst actions and prices: Yahoo Finance. The data itself is not included here.
• Logistic regression with balanced class weights; C = 10.0 chosen on validation 2024 (train < 2024). The released weights are refit on all years with that C.

Evaluation

Rolling-origin test: each year predicted by a model trained only on earlier years (2,816 announcements, 469 misses):

Model	Pooled AUC	ΔAUC vs history [95%]	Precision in top 10%
history only	0.655	—	34%
history + sector	0.670	+0.015 [-0.002, +0.033]	36%
this model	0.680	+0.025 [+0.004, +0.047]	36%

By test year:

Test year	Announcements / misses	history + sector	this model	Δ
2023	726 / 138	0.670	0.704	+0.034
2024	727 / 135	0.658	0.679	+0.021
2025-26	1,363 / 196	0.673	0.671	-0.002

Base rate 16.7%. On companies held out of training (5-fold by company): AUC 0.688, +0.020 [+0.010, +0.030] over history + sector. The analyst + price part of that gain (+0.019) is unchanged when each test row's own year is left out of training, so it is not a same-quarter market shock.

Limitations

• Modest skill. AUC ≈ 0.68 ranks misses better than chance, not reliably. Flagging the riskiest 10% finds misses at 2.2× the base rate.
• The most recent test period (2025–26) shows no gain over history + sector. With three test years we can't tell a change in markets from noise.
• Depends on Yahoo Finance at prediction time. Yahoo's analyst-action history is used as if it were complete and unrevised; that isn't guaranteed.
• The text adds little on its own; an LLM reading the same guidance passages did no better.
• Survivorship: companies are today's S&P 500 members (plus 49 large caps); REITs are excluded because their EPS labels mix FFO and GAAP bases.
• Probabilities come from class-weighted training, so they are not calibrated (they overstate the miss rate). Use them for ranking.
• Not investment advice. Educational project only.