Post-Earnings Drift, Decomposed: 81 Walks, a Predictability Spread That Was Market Exposure, and a Screen That Picks Better Announcements and a Worse Book

This is the second earnings-drift paper on this site. The first walked 220 sealed studies: eleven sector books at five holding periods under two anchors, the announcement and the quarter end. It found the drift alive in Technology and Industrials, reversed in Utilities, a short leg that died at the index level, and a surprise screen that added nothing to a book entered on time. It left one question open: whether those sector results were about the industry label or about the companies inside it, and it named a beta-neutral rerun as the check it had not run.

This paper takes both up. It cuts the same index by the company's own earnings history instead of its sector, and it runs the hedge. The first paper, with its frozen tables, its 220 cells and the register of every attempt: Post-earnings announcement drift 2006-2025: 220 registered sector walks, two entry anchors, long-short +2.2% a year at the announcement, -0.5% at formation.

What is bought, and when. The company reports. The number becomes public, after the bell counting as the next session. The position opens at that session's close and is held for exactly 5, 10, 21 or 42 trading days, then closed. One open position per name, each at a tenth of book equity, equal split past ten. Three books do this on the same announcements of the same cohort: the long-short book buys every qualifying beat and shorts every qualifying miss; arm A buys only the top-quintile beats; arm B buys every announcement with no screen at all. The first figure draws the three.

Every result in this paper comes from one rule set, the one the first paper registered for its announcement wave, run on nine populations instead of eleven sectors. Nothing about the entry, the holding, the costs or the floors changes between cells. What changes is which names are eligible.

The populations. At each window's anchor the point-in-time S&P 500 membership is scored on two properties, both knowable at that close. Earnings predictability is the standard deviation of the name's own last eight forecast errors (actual EPS minus consensus, at least four required) divided by its mean absolute estimate over the same quarters, so a two-dollar earner and a twenty-cent earner sit on one scale. This is the denominator of the SUE measure itself, read as a property of the company. Gross profitability is Novy-Marx's gross profit over total assets, taken from the platform's point-in-time fundamentals lane, which admits a filing only once its SEC acceptance date has passed. Each score is cut into nearest-rank terciles on the cross-section present at that anchor, and the cohort is one cell of the three-by-three split. A name missing either score is refused rather than filled. The cohort is fixed for the window; the announcements it makes over the following year are the events.

The surprise. SUE is actual EPS minus consensus over the spread of the name's own last eight forecast errors, at least four required, capped at ten either side, the same measure the first paper registered.

The bar. The long-short book buys a beat whose SUE sits at or above the 80th nearest-rank percentile of the whole index's trailing year of surprises and shorts a miss at or below the 20th, with the sign required to agree with the side and a zero surprise never entering. The thresholds come from the whole index, not the cohort, so the bar is identical in every cell and a cohort that clears it less often simply trades less. The long-only pair registers arm A, the same top-quintile rule without the shorts, against arm B, every announcement in the cohort. Costs are liquidity-tiered spreads by 63-day dollar volume, 1.5, 3 or 6 basis points one way at entry and at exit; shorts pay fifty basis points of borrow; dividends credit from the payment record. A window whose cohort has fewer than fifteen distinct announcing names, or in which no announcement clears the bar, is excluded and printed as excluded.

The yardstick. Every long-only cell is compared with the S&P 500 total return over the same sealed window, computed on the engine's own convention from SPY closes with cash distributions credited on their ex-dates, paired window by window, twenty windows.

The walk. Nine cohorts by four rungs by two book shapes is 72 registered studies in the grid, 81 with the hedged reruns, each walked through twenty sealed one-year windows anchored every January from 2006 to 2025: 1,440 windows in the grid, 1,620 with the reruns. The prospectus of each step is sealed before its window runs, and the family of 72 is declared on every page, so the deflation statistics answer to the whole search. The point-in-time index, the earnings table with its known-from dates and the liquidity tiers are the first paper's frozen tables; the fundamentals are the platform's own lane. Two defects were found in rehearsal and fixed before registration: the run summary dropped the cohort record, and the prospectus did not yet count the cohort book as a forward test. The walk itself ran without a failure.

What this design can and cannot say. It measures where the announcement reaction lives across two firm properties on one fixed rule. It does not tune the rule, does not vary the surprise measure, and does not pool across cohorts; a reader who wants the pooled index answer has the first paper.

The long-short grid is Table 1, and its rows are the first thing a reader meets: the steady row averages +3.0 a year net across its twelve cells, the middle row +2.8, the erratic row +1.1. The steady, strong corner climbs with the holding period, +2.1 at five days, +3.9 at ten, +4.8 at a month and +6.6 (±1.6) at two, positive in 15 of twenty windows at 42 days; the erratic, weak corner peaks at ten days at +2.1 and reverses to -3.0 (±3.8) at two months. On the same twenty windows the steady corner leads by +1.1, +1.8, +3.4 and +9.6 points at the four rungs, ahead in 11, 13, 14 and 15 of twenty; at 42 days the sign test puts the count at p = 0.02 (Table 6). Compounded across the stitched windows the steady, strong cell grows one dollar to 3.45x and the erratic, weak cell to 0.42x, against 7.80x for the index.

Arm A of the long-only pair is the long-short book's long leg by construction, the same entries at the same sizing without the shorts, so the long-short return minus arm A on the same windows is the short leg. At 42 days it loses in all nine cohorts: 4.0 points a year in the steady, strong corner, 8.4 in the erratic, weak corner, where the long leg alone returns +5.4 (Table 5, Figure 3). Shorting misses lost money in every population and lost most where earnings were least predictable, because clearing an index-wide bar from an erratic earner takes a wide miss, and wide misses move on the day and recover.

Nine more sealed studies rerun the 42-day long-short cells with a daily hedge: the book is charged its net exposure times the S&P 500 total return, debited on a net-long day in a rising market and credited on a net-short one. Table 5 prints the unhedged and hedged books on the same windows and their difference. The steady, strong book ran 38 percent net long; the hedge took 4.2 points a year (±1.7), leaving +2.4 (±1.5), positive in 12 of 20 windows. The erratic, weak book ran 31 percent net short; the hedge gave it 5.3 points, taking it from -3.0 to +2.3. The corner gap of +9.6 ±4.2 becomes +0.1 ±3.9, the steady side ahead in 8 of 20, and the hedged rows read steady +0.9, middle +1.4, erratic +1.9, with 7 of the nine hedged books positive and no ordering by predictability left. The 42-day spread between predictable and erratic earners in the long-short book was net exposure, 64 percent of the steady corner's return; what the hedge leaves is a small announcement premium in most cohorts, a point or two a year, that costs of 0.5 points already eat into.

Inside each cohort the top-quintile book trails the every-announcement book by 2.0 points a year on average, in 30 of 36 cells; in the steady, strong cohort at 21 days the gap is 5.7 (Table 4). The screen is not picking worse announcements. Per dollar deployed the screened entries earn more, +14.6 a year across the grid against +9.9 for the unscreened book (the grid means over the grid deployed shares), and at 42 days +12.9 against +13.2 in the steady, strong cohort. What the screen does is leave the book idle: a top-quintile bar on the whole index admits few names per cohort, so the screened book holds 75 percent of its capital in cash on an average day against 43 for the unscreened one, and idle cash earns nothing in this engine. Crediting the idle gap at the era's bill rate, about 1.5 percent a year, returns 0.5 of the 2.0 points; at today's rates it would return most of them. The exceptions sit in the erratic row at the short rungs, where the screen adds a point or two to the book as well.

The long books are reported on three bases because they do not agree, and against the index, which is the benchmark this family recorded (Tables 2 and 3). Every announcement of the steady, strong cohort held 42 days made +12.5 a year (±3.5) against +12.4 for the index on the same windows, ahead in 9 of twenty; compounded it grew one dollar to 8.76x against 7.80x; its pooled daily Sharpe was 0.68 against the index's 0.63. Across the 36 every-announcement cells 1 finished ahead of the index on the mean of windows and 3 compounded higher; the top-quintile book trails on both in every cell, with 82 percent of its capital deployed in the steady, strong cohort at 42 days and +12.9 earned on it. On row means the every-announcement book orders by predictability at 42 days, steady +10.5, middle +9.4, erratic +8.4, and on deployed capital +11.7, +10.2, +9.5; three of the nine cells break the pattern at that rung, the erratic and middle cell at +11.5 above steady and middle at +10.0 and steady and weak at +8.9. That ordering is a return spread between the cohorts as populations. At 42 days the book holds about 95 percent of the cohort with a lag, so what it measures is mostly the cohort. The comparison that would isolate the announcement, each cohort's book against the same cohort held outright over the same windows, is not in this family and is not claimed.

Predictability is measured on the same eight-quarter error history the SUE denominator uses, so with an index-wide bar a steady name enters on a small dollar surprise and an erratic name only on a large one; steady versus erratic is partly small versus large absolute surprise. Table 5 shows it in the entry counts: the steady, strong cohort made 100 long-short entries a window and the erratic, strong cohort 38. Arm B applies no surprise screen and is immune to the objection; its rows order the same way on pooled means, +7.4 against +3.6 (Table 7), and what that ordering is made of is the question the previous paragraph leaves open.

Profitability is the secondary axis. Across the long-short book the strong column averages +3.2, the weak +2.2 and the middle +1.5: not monotonic, and in the screened long book the middle tercile is again the weakest. Hedged, the middle column is the one negative column (-0.7 against +2.5 and +2.5). Profitability sharpens the steady row in the long books and ranks nothing on its own.

The cohorts are unequal because the two properties are correlated: steady earners are more often profitable and erratic earners more often weak, so the agreeing corners hold 57 and 69 names at a typical anchor and the disagreeing corners 29 and 29 (Table 5). The fifteen-name floor refused 27 windows in all, every one in the two small corners, and every one is printed. Beats dominate the steady cohorts and misses the erratic ones, which is where the net exposure in Table 5 comes from: 5 of the nine long-short books ran net long on average. Costs drag the long-short book 0.5 points a year on average.

The result is a decomposition, and it is not the one either literature asked for. Bhushan and Mendenhall put the drift where arbitrage is hard, Cao and Narayanamoorthy where the surprise is informative, and the unhedged long-short grid seemed to answer for the second: the steady corner kept going for two months and the erratic corner reversed. Nine sealed reruns with the market exposure charged take that answer away. The steady corner was net long and the erratic corner net short in a market that returned +12.4 a year, and once the exposure is charged the two corners sit +0.1 ±3.9 apart. What the announcement trade paid, hedged, is a point or two a year in most cohorts, before the question of which population earned it can be asked.

What a practitioner takes from it. A long-short drift book confined to predictable earners is not a predictability spread; sized to the hedged number it is a small premium that costs of 0.5 points a year already eat into, and the short leg lost in every cohort. The surprise screen belongs on the entry and not on the book: it raises the return per dollar deployed from +9.9 to +14.6 across the grid and lowers the book's return by 2.0 points because the capital it does not deploy earns nothing here; a book that keeps its idle capital in bills recovers 0.5 of those points in this era and most of them at today's rates. Long only, the announcement anchor bought little beyond the index: the best every-announcement book matched it on the mean of windows and beat it compounded, and the rest trailed on both bases.

Where it leaves the first paper. The sector grid said Technology and Industrials carried the announcement drift and Utilities reversed. Read through this grid, part of that was composition, sectors full of steady, profitable earners against a sector of regulated, low-surprise names, and part of it was the same net exposure this paper charges; the sector paper did not run the hedge, and its long-short cells should be read with this one's Table 5 beside them.

Where the ordering stands. The every-announcement books of the steady cohorts returned more than those of the erratic cohorts at every rung on pooled means, and three of nine cells break it at 42 days. At that rung the book holds most of the cohort with a lag, so the ordering is a return spread between predictable, profitable large caps and erratic, weak ones over 2006 to 2025, with an announcement label. Whether the announcement adds anything to owning the cohort is the comparison this family did not record, each cohort's book against the same cohort held outright on the same windows, and it is the check the next registration carries.

What a reader can check: every one of the 81 cells is a page on this platform with its sealed prospectus, its per-window reports, its exclusions printed, and the family of 81 in the footer. This page is the steady, strong long-short cell at 42 days. The first paper, its frozen tables and its 220 cells are at https://quanterlab.com/research/post-earnings-drift-by-sector-walked-220-times-under-two-anchors-dead-at-the-for.

One era: 2006 through 2025, one bull-heavy market regime in which the index returned +12.4 a year on these windows and compounded to 7.80x, twenty windows per cell, and the standard errors in the tables are across those twenty windows. The two small corners, steady and weak and erratic and strong, hold about 29 names at a typical anchor and lose one or two windows each to the fifteen-name floor; every exclusion is printed. Daily closes only: entry is the close of the first session the announcement is known, so the overnight gap and the first day's reaction are excluded by construction, which understates what faster execution captures. Costs are tiered spreads without market impact, fair for large-cap books of this size and untested at scale; borrow is a flat fifty basis points. One surprise measure, one predictability measure and one profitability measure by design; other definitions were not run and are not claimed. The consensus history is one vendor's record.

The long books have no cohort-held benchmark in this family: the recorded benchmark is the index, and at 42 days the every-announcement book holds about 95 percent of the cohort with a lag, so its ordering across cohorts is a spread between the cohorts as populations and the announcement's own contribution is not separated. The long-short book floats its legs; the leg decomposition in Table 5 assumes the two legs are struck on the same equity, which holds to the sizing rule's approximation. The hedge charges net dollar exposure against the index, not an estimated beta per name, so a book of high-beta names keeps some market in it after the charge; it was run at the 42-day rung only. Idle cash earns nothing in this engine, which understates every long book against an investor who parks cash in bills, by about 0.5 points a year on the idle gap between the two long books in this era.

Where the story stops: the grid is measured, and nothing here has been deployed.

1. Ball, R., & Brown, P. (1968). An empirical evaluation of accounting income numbers. Journal of Accounting Research, 6(2), 159-178. (First documentation of drift in the direction of the earnings news after the announcement.)
2. Bernard, V. L., & Thomas, J. K. (1989). Post-earnings-announcement drift: delayed price response or risk premium? Journal of Accounting Research, 27, 1-36. (The drift concentrated in the weeks after the announcement.)
3. Bhushan, R. (1994). An informational efficiency perspective on the post-earnings announcement drift. Journal of Accounting and Economics, 18(1), 45-65. (The drift larger where the cost of trading against it is higher.)
4. Mendenhall, R. R. (2004). Arbitrage risk and post-earnings-announcement drift. Journal of Business, 77(4), 875-894. (The drift concentrated where idiosyncratic risk deters arbitrage.)
5. Zhang, X. F. (2006). Information uncertainty and stock returns. Journal of Finance, 61(1), 105-137. (Greater uncertainty, larger drift after news.)
6. Chordia, T., Goyal, A., Sadka, G., Sadka, R., & Shivakumar, L. (2009). Liquidity and the post-earnings-announcement drift. Financial Analysts Journal, 65(4), 18-32. (The drift concentrated in illiquid names once trading costs are charged.)
7. Cao, S. S., & Narayanamoorthy, G. S. (2012). Earnings volatility, post-earnings announcement drift, and trading frictions. Journal of Accounting Research, 50(1), 41-74. (Lower earnings volatility, more informative surprises, larger drift.)
8. Novy-Marx, R. (2013). The other side of value: the gross profitability premium. Journal of Financial Economics, 108(1), 1-28. (Gross profit over assets as the quality measure.)
9. Martineau, C. (2021). Rest in peace post-earnings announcement drift. Critical Finance Review (SSRN 3111607). (Drift in large caps shrank toward zero on standard designs.)
10. QuanterLab (2026). Post-earnings drift by sector, walked 220 times under two anchors. https://quanterlab.com/research/post-earnings-drift-by-sector-walked-220-times-under-two-anchors-dead-at-the-for (The first paper in this series: the anchor result and the frozen tables this grid reuses.)