Quality
Is this company worth the event?
The method
A market inefficiency documented for 50 years.
Post-earnings drift, modeled with the same three-layer architecture a quantitative fund would build — and shipped ready to use.
Post-earnings drift, in plain terms
After an earnings announcement, the price tends to keep drifting in the direction of the surprise — for up to 60 days. The anomaly has been documented across five decades of academic literature and has survived index changes, regime shifts, and the rise of algorithmic trading. The mechanism is underreaction: the market prices the reported number almost instantly, but takes weeks to fully incorporate its implications. That lag between information and its full price is the edge.
EarningSpy does not predict the surprise — beats, misses, the gap to consensus. What we model is the probability of a market-risk-adjusted abnormal return in the 60 days following the announcement, and its direction. In 82% of historical cases that move persists.
We cover the full cycle — pre and post-announcement. The signal runs through cascade inference across multiple models, built with the rigor of a hedge fund. What changes is who gets to use them.
Anomaly
Will drift appear at all?
Direction
Which way will it go?
L1
Quality — is this company worth the event?
Risk-adjusted metrics retuned to the earnings window — the same recalibration event-driven desks run, with PCA compressing the result into a single quality read.
Every earnings model starts with a filter: which companies are worth modeling in the first place? Retail platforms answer that with off-the-shelf metrics — a Sharpe ratio computed over months of daily returns, a Sortino built on a generic universe. Useful, but blunt: those numbers describe past behavior in general, not behavior around an earnings event. Event-driven desks recompute them on the window that matters; outside of them, almost no one does.
Layer 1 runs that recalibration. Sortino, Calmar, Sharpe and Information Ratio are recomputed using returns observed in the days surrounding past earnings prints — so the resulting numbers describe how a company has behaved during the event, not on a random Tuesday.
The four calibrated metrics
- Sortino downside-aware
- Calmar drawdown-aware
- Sharpe volatility-aware
- Information Ratio benchmark-aware
Four metrics looking at the same window is informative but noisy — they share a lot of signal and they disagree at the edges. To get one clean read of quality, the layer runs a principal component analysis across them. PCA — a workhorse of quantitative finance for decades — finds the axes along which the four metrics vary together and compresses them into orthogonal components. The leading component captures most of the shared signal: that single number becomes the layer's quality read.
The output of L1 is not a buy/sell signal. It's a gate — the same kind of risk filter a fund applies before any position-level model gets to run.
L2
Anomaly — will drift appear at all?
A machine-learning classifier predicts the presence of post-earnings drift, independent of its direction — the first stage of a two-stage architecture used in quantitative event-driven strategies.
Not every earning produces a drift. Some announcements are absorbed by the market within hours; others trigger a multi-week trend in the direction of the surprise. Trading a drift that never materializes is the most common way to lose money on this anomaly — and the most common reason naive earnings strategies fail.
Layer 2 answers the binary question first — will drift appear, yes or no — before any direction is considered. Splitting presence from direction is deliberate, and standard practice in production-grade quantitative models: they are different signals, with different drivers, and a model that tries to solve both at once tends to do neither well.
A machine-learning classifier — trained quarter by quarter on historical earnings and their realized drifts — returns a probability that drift will appear in the post-announcement window. High confidence is the gate to L3. What was built and tuned inside event-driven funds for years is what runs here, every quarter, applied to the broad universe.
L3
Direction — which way will it go?
A second classifier — trained separately, applied only where L2 flagged drift — predicts the sign of the move.
Direction is the harder problem. A positive surprise does not guarantee an upward drift; a miss does not guarantee a downward one. Because direction depends on different signals than presence, it gets its own model — same family as L2, but trained on a different target.
Cascade inference
L3 only runs on the events L2 flagged. By filtering out cases where drift was unlikely in the first place, the direction model trains and operates on a cleaner distribution — events where something real is happening. Cascade inference is standard architecture inside production ML pipelines at quantitative funds; applied here to the earnings problem, it is what lifts the precision of the second model.
The combined signal — anomaly present, with a direction attached — is what the next layers turn into a trade.
Two ways to use the model
Stop at L2 — trade volatility
For market-makers and volatility-focused desks who already run direction-neutral structures around events: the anomaly read alone is enough.
Go to L3 — trade direction
For directional funds and sophisticated retail traders: the next two layers — risk management and position sizing — turn it into an actionable long or short.
Sizing
How much to risk?
The same Kelly-based sizing that runs inside quantitative funds — now wired into the model's own output.
Inside quantitative funds, position sizing is not a discretionary call. It runs on the Kelly criterion — a formula that translates a probability of winning and a payoff ratio into the bet size that maximizes long-run growth. The math has been in the playbook of professional capital allocators for decades. What has been missing for everyone else is the rest of the stack: a model accurate enough to feed it, and infrastructure that wires the two together.
EarningSpy closes that gap. The directional probability and the expected risk-reward from the previous layers feed straight into the formula. Higher confidence with an asymmetric payoff justifies a larger position; lower confidence pulls the size down. Sizing becomes a consequence of the model, not a separate guess.
Why fractional Kelly
Full Kelly is mathematically optimal but practically punishing — its drawdowns are larger than any real operator wants to live through. The same funds that pioneered Kelly almost never run it at full size. Neither do we: EarningSpy applies a fractional Kelly, typically below half, that trades a bit of theoretical growth for materially smaller drawdowns. Aggressive enough to compound, conservative enough to survive.
For a sophisticated retail trader, that means access to the same risk-of-ruin discipline that allocates capital at a quant fund. For a smaller fund, it means a sizing layer that would otherwise require a dedicated quant team to build. The toolkit is the same; the gate is who gets to use it.
Where
The anomaly doesn't live everywhere.
PEAD concentrates in specific sectors and clusters — and the model tells you which ones, quarter by quarter.
Five decades of academic literature have documented the same thing: post-earnings drift is uneven across the market. It is stronger in some sectors than others, more pronounced in companies of a certain size and analyst coverage, more pronounced still in clusters of companies that share factor exposures. A model that ignores where the anomaly lives spends most of its time looking in the wrong places.
EarningSpy maps the universe into those clusters and tracks where the anomaly is currently active. The result is a curated watchlist — not the full market, but the subset where the model has historical evidence and current signal. It is the same kind of universe-construction work a quantitative fund does before any trade is sized.
Outside that subset, the model stays quiet. That restraint is the point.
Recalibration
A model that only gets better with time.
Every quarter, the entire pipeline retrains on the most recent earnings cycle — the same operating cadence a quantitative desk runs.
Markets change. The factors that drove drift two years ago are not the same as the ones driving it today, and they will not be the same two years from now. A model that doesn't retrain quietly decays — and most retail-facing tools never retrain at all.
EarningSpy recalibrates end to end every quarter. The risk metrics, the PCA components, both classifiers, the sector and cluster maps — all of them are refit on the latest realized earnings and their drifts. The model that runs this quarter is not the one that ran last quarter. It is sharper, because there is one more cycle of evidence to learn from.
The toolkit. Ready to operate.
EarningSpy is still being built. Leave your email and be among the first to put this method to work on the next earnings cycle.