The Reversal Index
I developed a Reversal Index (Ri) to compare analysts' forecast accuracy when earnings change direction with their accuracy in other quarters. The course project added Ri to an analyst-ranking model and tested a trading strategy. This page shows those original results alongside a later walk-forward validation.
Why measure reversal accuracy
The project asks whether analysts differ in their ability to forecast a change in earnings direction.
Analysts often disagree. I wanted to know whose forecasts deserve more weight, particularly around a change in earnings direction.
IBES collects every sell-side analyst's EPS forecast. Averaging them gives the market consensus that investors, traders and the financial press quote. Shoorvazi (2025) improved on the plain average by scoring each analyst on accuracy, information diversity and consistency, then averaging only the top 25%.
That framework does not score reversal quarters separately. I wanted to test whether an analyst's relative accuracy changes when EPS reverses direction, and whether that distinction could improve the consensus forecast.
What counts as a reversal
A mechanical rule with two conditions: the direction flips, and the move is large enough not to be noise.
- 01Direction flips:
sign(ΔEPSt) ≠ sign(ΔEPSt−1) - 02Move is material:
|ΔEPSt| / |EPSt−1| > 10%
Applied to Apple, 65 of 156 quarters (42%) qualify. Reversals are not rare events.
Quarterly EPS has a calendar. A holiday-quarter spike followed by a normal quarter looks like a "reversal" every year. Summing to annual EPS strips that seasonality out. Try it below.
The Reversal Index
Forecast accuracy in reversal quarters relative to accuracy in other quarters.
Higher relative forecast accuracy in reversal quarters.
The same relative forecast accuracy in both groups of quarters.
Lower relative forecast accuracy in reversal quarters.
From analyst scores to a trade
Ri becomes the fourth factor in Shoorvazi's scoring model. The top-scoring analysts form a refined consensus, and the gap between it and the market consensus becomes a trading signal.
Average of 1 − |forecast − actual| / |actual|, clipped to [0, 1].
How far the analyst's forecast typically sits from the crowd.
Inverse-variance share of the analyst's forecast errors over time.
Reversal-quarter accuracy divided by normal-quarter accuracy.
Example: Apple earnings, 29 Oct 2020
Score analysts
Grid-search w₁…w₄ in 0.2 steps per stock; score every analyst.
Keep the top 25%
Rank by score and keep the top quartile.
Refined consensus
Average the selected analysts' EPS forecasts.
$0.837Compare to market
Market consensus averages everyone.
$0.781Signal
(refined − market) / |market|
+7.2%Sort & hold
Each quarter, split stocks into terciles by signal; hold 60 trading days.
- Signal
- +7.2%
- Actual EPS
- $0.73
- 60-day return
- +29.0%
- Tercile
- Top
Original results and walk-forward tests
I first evaluated the model in-sample. A later implementation rebuilt the pipeline walk-forward, using only information available at each point in time. Switch between the two below; every chart on the page follows the switch.
In-sample · as submitted
Weights are grid-searched on all of a stock's history, including the quarters being scored. This is the version in the final report and deck.
Walk-forward · audited
Features, Ri and weights use only prior periods. Forecasts must be dated before the announcement. The 3-factor baseline is optimised independently.
Phase 3 · Forecast comparison — 15-stock pilot
Reversal-quarter forecast error: 4-factor vs 3-factor
% reduction in MAE · right is betterData table
Phase 4 · Annual and quarterly Ri
Trading backtest
68 stocks, 2005–2024, 4,837 earnings announcements. The top tercile is held long; the bottom tercile is also held long as a control; the long-short is top minus bottom, dollar-neutral.
Growth of $1 · 60-day holding window
log scale60-day results
Sharpe ratio by holding horizon
trading daysLong-short return by sub-period
annualised · 60-daySubsequent validation
The original course results compared with a later walk-forward implementation.
| Claim | In-sample | Walk-forward | Verdict |
|---|
What the walk-forward rebuild changed
- No look-aheadAnalyst features, Ri and weights are estimated only on periods before each test quarter.
- Forecast timingOnly forecasts activated before the announcement date count.
- Fair baselineThe 3-factor model is optimised on its own, not derived from 4-factor weights.
- Overlapping positionsDaily accounting treats overlapping 60-day windows as real concurrent positions.
- True dollar-neutralLong-short is long basket minus short basket; cash when one side is empty.
Research conclusions
Differences between analysts
The original project proposed Ri as a measure of reversal-quarter accuracy relative to normal quarters, then added it to an analyst-ranking model. The annual-versus-quarterly comparison examined how seasonal earnings affect that measure.
Weight selection and overfitting
The original pilot reported about 12% lower forecast error after choosing weights on the same data. The later walk-forward comparison did not reproduce that improvement. The two results answer different questions and are shown separately above.
Long-only and long-short results
The course presentation emphasized the top tercile's performance and the decline in long-short returns after 2011. The bottom tercile also beat SPY. That comparison, together with the later validation, limits how much of the long-only return can be attributed to the ranking signal.