Goldman Sachs / Morgan Stanley "Reverse Indicator" Test

Verdict: No, not "mostly against" — the full-sample contrarian rate runs 36.6% to 49.2% across all eight horizons, never sustainably above 50%; at +1Y it falls to 38.6% and the direction-aligned return turns positive at +5.83%.

76 major revisions, 464 valid observations across eight horizons. The contrarian rate oscillates between 36.6% and 49.2% at short horizons and falls to 38.6% at +1Y; direction-aligned mean returns are +0.09% / +0.16% / -0.42% / +0.85% in the first four short windows and +0.53% / +1.35% / +3.60% / +5.83% in the last four. "GS/MS are reverse indicators" therefore fails as a general claim and survives only as a conditional claim about horizon, action, firm and asset class: 57.5% after upgrades, Morgan Stanley at +1D, and single-stock calls at 75.0%. Splitting into layers is presentation only — the headline contrarian rate keeps the full 76-event denominator.

38.6%

+1Y contrarian rate

n=44, agree 27 / contra 17

57.5%

+2W after upgrades

n=40, strongest conditional signal

58.8%

Morgan Stanley, +1D

n=34; GS 35.7% same window

+5.83%

+1Y aligned return

holding in the direction of the call

1. Direct answer: contrarian rate by horizon (incl. 3W / 6M / 1Y)

Definitions: a fall after an upgrade, or a rise after a downgrade = contrarian. Rate = contrarian count / valid observations (zero returns and unmatured windows excluded). Direction-aligned return = +r for upgrades, −r for downgrades; >0 means the call worked, <0 means the market went against it. Eight identification horizons: +1 / +5 / +10 / +15 / +20 / +60 / +120 / +250 trading days ≈ 1D / 1W / 2W / 3W / 1M / 1Q / 6M / 1Y; +15 (3W) is newly added on the same basis as the other seven. Prices: Wind Financial MCP daily closes, 2006-01-03 to 2026-11-06. Events come from public media and secondary accounts of research notes; dates in Section 5.

Contrarian rate by group (eight horizons)

Dashed line = 50% random benchmark. The full sample hugs 50% at short horizons and drops to 38.6% at +1Y; differences between groups exceed differences between horizons → the effect is both a horizon problem and an event-type problem. Group sizes: GS 42 / MS 34 / upgrades 43 / downgrades 33.

How to read this: the contrarian rate answers was the direction right?; the aligned return answers was the magnitude worth it?. Full-sample rates are 46.1% / 36.6% / 49.2% / 39.0% / 42.9% / 42.0% / 42.6% / 38.6% across the eight horizons — never sustainably above 50% — while the aligned mean climbs from +0.09% to +5.83%. The reverse pattern is not a majority phenomenon but a conditional one amplified by a few events, upgrades and short horizons.

2. Curve comparisons

Every curve uses "trading days after the event" on the x-axis, with the event-day base return set to 0. Curves are cross-sectional averages; days with fewer than 3 valid events are left blank (no interpolation). Above zero = market moved with the call; below zero = against it.

Fig. 1 · Direction-aligned return path: all / upgrades / downgrades (+1 to +60)

All three lines spend most of the first quarter below zero → holding in the direction of the call lost money on average. Upgrades sit below downgrades: upgrades get punished more often. Samples: 76 / 43 / 33.

Fig. 2 · Raw price path: upgrades vs downgrades

No direction alignment — actual returns. "Upgrades trending down, downgrades trending up" is the visual signature of a reverse indicator. Note both lines often move together (shared market beta), which is exactly the lack of market-neutralisation discussed in Section 6.

Fig. 3 · By period: 2006–2016 vs 2016–2026

P1 (10 events) and P2 (66 events) diverge clearly; +1Q aligned returns are +6.78% / +0.00%. The reverse pattern is more pronounced in the last decade. Sell-side coverage, information diffusion and passive flows are plausible contributors, but this sample cannot identify causality.

Fig. 4 · By firm: Goldman Sachs vs Morgan Stanley

GS (42 events) at +1D shows 35.7%; MS (34 events) shows 58.8%. Lumping the two together as "GS/MS reverse indicators" hides this gap, though MS is more concentrated in HK / China assets, so firm and market effects are confounded.

Fig. 5 · By asset class

Under the legacy cat grouping: global indices 29, commodities 16, key stocks & sectors 25, sectors 6. That grouping mixes single stocks with sectors, so the next section re-cuts it into four disjoint classes (indices / sectors / commodities / stocks) and reports a contrarian rate for every partition.

Fig. 6 · Long horizon: does the reversal arrive late? (+1 to +250)

Vertical lines mark +60 (1Q), +120 (6M), +250 (1Y). The classic case P01 (P01, 2008-05-06) peaks at 74.55% on T+204 — a one-quarter window misses it entirely. Yet the full sample is only 38.6% contrarian at +1Y with +5.83% aligned return, so a "late reversal" is an exception, not a rule.

Fig. 7 · Actual price paths of landmark events (event day = 0)

Eight representative events, x-axis to +250 trading days; colours distinguish events only. P15 (2022-10-26, MS cut HSI to 16900) subsequently rose +47%; P06 (2013-04-10, GS advised shorting gold) fell −17% — textbook contrarian and agree cases respectively. P01 shows the "invisible at one quarter, obvious at six months" pattern.

3. Four conditional signals that only appear after splitting the sample

① ActionUpgrades reverse more
+2W 57.5% vs downgrades 33.3%

›

② FirmMS contra, GS near neutral
+1D 58.8% vs 35.7%

›

③ Asset classStocks & commodities diverge most
Stocks +2W 75.0% / indices +1W 27.6%

›

④ HorizonThe effect expires
+1Y contra rate falls to 38.6%

3.1 By action / firm / period (eight horizons)

3.2 Legacy asset grouping (compare with the new four classes)

3.3 By instrument

The single most important split: direction-aligned returns after upgrades are -0.53% / -0.46% / -1.33% / +1.02% / +1.16% / +1.76% / +5.20% over the first six windows, turning positive only at +1Y; downgrades are mostly positive, reaching +2.49% at +1Y. Major upgrades tend to arrive when price has already reached a key level and the narrative is loudest — trend confirmation rather than a leading warning. This is correlation, not causation.

3b. By asset class: indices / sectors / commodities / stocks

The 76 events are cut into four disjoint classes: 29 equity-index, 6 sector, 16 commodity and 25 single-stock events, summing to 76 — exactly the full sample. Each layer reports its own contrarian rate, aligned return and contrarian magnitude across all eight horizons. Layering is presentation only: the headline contrarian rate keeps the full 76-event denominator, with each window using that window's matured, non-zero observations.

Denominator unchanged: the four classes are mutually exclusive and sum to the full sample; layering is a slice, not a re-weighting. The headline rate always uses the full sample, each horizon dividing by its own matured, non-zero observations. Do not average the four layer rates to reconstruct the headline — the TOTAL row is the only headline basis.

3b.1 Four asset classes · contrarian rate by horizon (last row = full-sample headline)

Fig. 7b · Direction-aligned paths by asset class (T+1 to T+250)

Each line is a layer's cross-sectional mean out to T+250; above zero = moved with the call, below = against it. Indices are lowest at +1W (27.6%), stocks highest at +2W (75.0%), commodities reach 53.8% at +1Q — between-layer differences exceed between-horizon differences. The sector layer has only 6 events with unmatured long windows, so the trailing blanks mean insufficient sample, not zero return.

3b.2 Contrarian rate for every partition

Reading the partition table: each row is one slice and all eight columns are that slice's contrarian rate; below 50% means the market mostly moved with the call. Event counts appear in 3b.1 / 3b.3. Expanded coverage (one note counted once per market it touches) is not independent and is shown for breadth only — it never enters the headline rate. 6 of the 6 sector events were split from shared source notes; robustness recalculations are in Section 4.

3b.3 Watchlist · event count and contrarian rate by group

3b.4 US top-30 by market cap · integration status

3b.5 Primary market vs expanded coverage

Prices are integrated for every watchlist group, but for Silver, Major HK-listed companies no verifiable GS/MS major upgrade or downgrade was found in the study window, so their event counts are honestly left at zero rather than fabricated. 9 of the US top-30 names are integrated; the remainder are blocked by the monthly batch quota and the daily tracker will backfill them and fold any new revisions into the statistics.

4. Robustness checks

Seven recalculations address clustering of recent events, dilution by unmatured windows, a single extreme case, low-quality sources and shared-source sector splits. Each cell shows the contrarian rate with the direction-aligned mean return below it, across all eight horizons. These are robustness checks only; the headline denominator is unchanged.

· Excluding events after 2026-09 and keeping only +1Q-matured events give near-identical results → unmatured recent events are not manufacturing a false reversal.
· Tier-A sources only (45 events) show +1M 46.4% and +1Q 45.5% contrarian rates with +3.37% aligned return → weaker sources dilute the signal, though tier-A events cluster in the last decade and are confounded with the period effect.
· Dropping P15 moves the +1Q aligned return from +1.35% to +2.35% → a single case visibly shifts the long-window mean, which should be treated as not robust. Excluding an event is a robustness check, not a claim that it is noise — extreme reversals are precisely where contrarian P&L comes from; see the magnitude section.

4.1 The two long identification horizons in isolation

+6M contrarian rate 42.6% (n=47); +1Y 38.6% (n=44, agree 27 / contra 17), aligned returns +3.60% / +5.83%. Neither long horizon makes the reversal more common. But events matured enough for +250 stop around mid-2025, so the long-horizon sample is structurally tilted toward early events and cannot separate horizon length from period.

Special topic · Contrarian magnitude: the biggest reversal is not noise

Every table so far answers was the direction right? (a count), not how far did it go the other way?. We add contrarian magnitude CM = max(0, −aligned return): zero when the call agrees, otherwise equal to the loss from following the note and to the gain from fading it. The contrarian rate stays near 38.6%, yet mean CM rises from 1.03% at +1D to 8.61% at +1Y (about 8.4x), and the top three events supply 29.1%–46.4% of total CM in every window. The largest reversal is therefore not noise but the main source of contrarian P&L.

Topic 1 · Magnitude and concentration across eight horizons

"Aligned all / ex-max" shows the leverage of extreme cases: +1Q moves from +1.35% to +2.35%, +1Y from +5.83% to +7.57%. The agreement result is built up from the majority of events where no large reversal occurred, while contrarian gains are highly concentrated in a handful of events. The largest single-window magnitude is P01 at +1Y (+69.05%); the largest peak event is P01 (T+204, +74.55%).

Fig. 8 · Mean contrarian magnitude, day by day (T+1 to T+250)

Y-axis is mean CM (%): higher means the market moved further against the call. Roughly +2.47% at T+20, +4.83% at T+60, +7.18% at T+120 and +9.78% at T+200. The reversal is not a one-day gap but accumulates over time, peaking between six months and a year. Upgrades sit above downgrades, consistent with upgrades being punished more often.

Topic 2 · Largest contrarian events (ranked by peak CM within T+250)

Peak day = the trading day when that event's CM was largest. 4 of the top 12 involve crude or gold: directional commodity calls are the main source of extreme reversals. P01 peaks at +74.55% on T+204. The pre-20D / pre-60D columns identify calls made after a run-up or a sell-off; source grades are explained in Section 5.

Fig. 9 · Magnitude paths of the five largest reversals

Each line is one event's cumulative CM (%) out to T+250. Large reversals come in two shapes: one spikes before T+60 (a sentiment reversal), the other accelerates only after T+150 (cycle peaks / fundamental turns). This is why a +1Q window misses the biggest cases.

Topic 3 · When is the reversal largest?

Two dimensions give an actionable answer. (i) Timing: peak days are bi-modal — fast reversals shortly after the note and large late reversals near T+200. (ii) Pre-event momentum: calls issued after an extreme 20-day move in either direction produce the largest reversals; range-bound tape produces the mildest. Reversals are largest when the note arrives after an extreme move. This is correlation, not causation, on 76 events — not a trading rule.

Daily tracking · GS / MS upgrades & downgrades

Each day we scan for newly published upgrades or downgrades from both firms on major global indices, commodities and key stocks / sectors (targets, ratings, price forecasts, outlooks), de-duplicate by date ±1 day + firm + instrument + direction, append to daily-tracker/events.jsonl, recompute the statistics, rebuild both editions and upload the refreshed deliverables.

Tracking since 2026-11-07 · Total logged 0 entries · Last update — · daily scan → de-dup → recompute → rebuild both editions → auto upload

Date	Firm	Action	Instrument	Change summary	Source	Grade	Status
Tracking is live; waiting for the first upgrade / downgrade entry.

Only Goldman Sachs and Morgan Stanley are tracked; JPMorgan, Citi, UBS and others are excluded. Entries without an exact date, without a tradeable instrument, or based on unverifiable second-hand accounts are logged but excluded from the backtest. Unmatured horizons stay blank (no interpolation). Source grades A / B are defined in section 5.1.

5. Event detail & return matrix

Actual returns (%) for 76 events across eight horizons. Default: green = up, red = down (international convention); switch or customise via Colour settings. Cell shading follows direction alignment: light blue = agree, light orange = contrarian, grey = unmatured / zero. Prices are Wind Financial MCP daily closes. Click any row to open that event's daily path in three views (actual / aligned / contrarian magnitude).

5.1 Sources & research-note summaries

Source grades: A authoritative financial media / institutional public abstract with an explicit date; B secondary account of a research note, original to be verified. This event list is not a primary research-note database; grade-B dates and magnitudes may be off by ±1–2 days, the largest input error in this study.

6. Methodology, source quality & limitations

6.1 Calculation conventions

Item	Convention
Firm scope	Strictly Goldman Sachs and Morgan Stanley. JPMorgan, Citi, UBS, BofA, Jefferies etc. are excluded even when named in the same article.
Event definition	A major upgrade or downgrade of an index target / rating, a commodity price forecast, a sector view, or a stock rating / target price, with a verifiable public date. No exact date, no tradable instrument, no price series → excluded.
Asset classes	Four disjoint classes: indices / sectors / commodities / single stocks, summing to the full sample. Layering is presentation only and does not change the headline denominator; sector calls use index proxies and are flagged as such.
Instrument mapping	Index targets → CSI 300 / Hang Seng / Hang Seng Tech / S&P 500 / Nasdaq 100; gold / silver / oil → spot gold / spot silver / Brent; sectors → matching CSI or thematic indices; stocks → adjusted A-share closes and explicit HK / US tickers. MSCI China and HSCEI are proxied by the corresponding market index.
Base price	Event-day close; for a non-trading day, the most recent prior trading-day close.
Identification horizons	Eight: +1 / +5 / +10 / +15 / +20 / +60 / +120 / +250 trading days ≈ 1D / 1W / 2W / 3W / 1M / 1Q / 6M / 1Y. Counted in actual trading days after holiday placeholders were removed.
Return	r = close on the N-th trading day / base close − 1; unmatured windows return null with no interpolation.
Agree / contrarian	Rise after upgrade / fall after downgrade = agree; fall after upgrade / rise after downgrade = contrarian; r=0 is flat.

6.2 Price-data cleaning audit

Early Wind series contain holiday placeholder rows (open = high = low = close, or duplicate dates). Left in, they shift every window by counting non-trading days. 275 rows were removed across 33 instruments. The v2 cleaning rule was corrected: a row is a placeholder only when all four prices are equal and the close equals the previous close; four-equal-price rows whose close changes are genuine trading days (common in early gold / crude / semiconductor index data) and are kept. Limit-up "one-price" days in individual stocks are kept.

6.3 Limitations that matter

Not excess return. Verdicts use absolute moves with no market beta or sector benchmark removed. Broad risk-asset sell-offs depress several upgrade events simultaneously; a strict event study should use benchmark-relative abnormal returns.
Observations are not independent. Repeated revisions on the same instrument create overlapping windows, and 6 sector observations were split from the same source notes (B* / U*). Section 4 provides "shared-source removed" and "one per 30-day cluster" recalculations. Effective independent sample size is well below 76.
Small sector / stock layers. Only 6 sector events (2 valid at +1Y) and 25 stock events, mostly from 2025–2026 with largely unmatured long windows. Their contrarian rates are descriptive only, not inferential.
Sector calls require index proxies. Constituents drift and the note's basket does not map one-to-one onto the index; B06 uses an A-share semiconductor index for a US semiconductor call and is a transmission observation only. 6 events carry a proxy flag.
Coverage counts are not publication counts. A single China-assets note often sets CSI 300 / Hang Seng / MSCI China / HSCEI targets at once; this study takes one tradeable proxy. HK therefore shows 4 events on the primary basis and 16 on the expanded basis — a proxy-mapping artefact, not fewer HK views.
Only 9 of the US top-30 by market cap are integrated. The rest are blocked by the monthly batch quota, and no verifiable GS/MS major rating / target change on them surfaced in public reporting over the last decade. No events were fabricated to fill the gap; the daily tracker will backfill after the quota refreshes. Watchlist groups currently at zero events are shown as zero.
Selection bias. "Major revisions" are defined by media availability, and the press disproportionately revisits forecasts later proven wrong — likely biasing the contrarian rate upward.
Targets are not trading signals. A 12-month target or rating is a relative-return framework, a different dimension from a +1 to +60 trading-day path; the +1Y horizon is actually closer to the sell-side's own time scale.
Input error. Grade-B dates and magnitudes need verification against the original notes; this is not a primary research-note database. Querying Wind by the Chinese name for "Intel" once fuzzy-matched to the Nasdaq Composite; explicit tickers (INTC.O etc.) were used to re-pull those series.

6.4 US & HK coverage (now in the statistics)

US single stocks are now in the statistics: querying Wind with explicit tickers (NVDA.O / AAPL.O / MSFT.O / INTC.O etc.) returns genuine single-stock daily closes, fixing the earlier Chinese-name fuzzy match. There are 32 US events in total, 17 of them on US top-30 names. HK / offshore China assets are reported on both a primary and an expanded basis (see 6.3 and the layer section). Cases previously listed as a qualitative appendix (Adobe / Intuit / China CXO / Arm) remain excluded because no verifiable exact publication date or usable price series exists.

6.5 Reproducible files & deployment layout

Price library market-data.json (33 instruments, cleaned 2006–2026 daily closes) → gs-ms-event-study-20y.js → study-result-20y.json → build_report.py compiles the static pages. Edit the event list and re-run to recompute every number, curve and layer table. Inside site/, index.html (Chinese, site root) and en/index.html (English, /en/) link to each other with relative paths; all data and the chart library are inlined — zero external requests, works fully offline. Live at contrarianindicator.net (Chinese at /, English at /en/), served by Caddy on a Tencent Cloud Lighthouse instance (Frankfurt) with automatic HTTPS; the 08:45 daily tracker rebuilds and syncs both editions.

Section 7 · Large guidance revisions and publication timing (second research line)

Upgrades / downgrades also have a meaning independent of sell-side research: a company's own earnings pre-announcement, flash report or guidance being revised sharply up or down, and the timing chosen to publish those numbers. This section uses twenty years (2006-2026) of structured A-share pre-announcement fields for market-wide statistics — the source is corporate disclosure, not broker research — with a +/-50% threshold on forecast net-profit change. All post-event returns are recomputed from the in-house 2006-2026 daily closes on the same eight horizons.

Figure 8 · A-share large earnings-guidance revisions, 2006-2025

Across two decades the market logged 9713 large upward (upper-bound forecast growth >50%) and 9053 large downward (profit warning / first loss, change < -50%) pre-announcements. Upward revisions peaked in 2021 (962) and downward in 2024 (1043); downward calls have outnumbered upward ones since 2022, tracking the earnings cycle. Disclosure clusters in the mandatory January / April / July / October windows. A-share pre-announcements cluster in the mandatory windows of January (annual), April, July (interim) and October (Q3), with January the clear peak. The month-level query uses a slightly different filter from the annual series, so monthly figures are used only for timing structure, not absolute comparison.

7.1 Twenty-year market-wide counts (corporate disclosure)

7.2 Publication timing vs contrarian rate (76 main-line events)

Each main-line event is bucketed by the percentile of its event-day close within the trailing 250 trading days. High (>=80th percentile): n=41, with contrarian rates of 56.8% at +2W and 50.0% at +1M, both above the full-sample 49.2% / 42.9%. Low (<20th percentile, n=15) rises above 55% only from +1M onward. Timing is an informative condition: optimistic calls published at price highs are punished more often, but the effect migrates across horizons and is not an unconditional contrarian signal.

7.3 Signal type: target/rating changes vs large earnings-forecast revisions

7.4 Guidance / estimate event pool (extended daily at 08:45)

Scope and caveats: 2026 is incomplete (the Wind annual series ends at 2025) and is filled by the daily tracker. A Wind natural-language aggregate of post-announcement 20/60-day returns returned a mis-matched static cross-section (identical values for the same stock across years) and was discarded. All post-event returns here are recomputed trade-by-trade with the in-house daily-close engine.
Engine: Company guidance events do not rely on sell-side research: the disclosure date is the event date and the threshold is a +/-50% change in forecast net profit. Post-event returns are recomputed from the in-house 2006-2026 daily closes using the same eight horizons. The pool is a growing starting set, extended by the 08:45 daily tracker.
The pool currently holds 2 entries (0 company guidance / 2 estimate revisions). The retro-added case: on 2026-07-17/18 Goldman Sachs and Morgan Stanley lifted Innolight's target price by 117% and its 2027 net-profit forecast by 180% on the day the stock plunged from its peak. Both were published at the 81.2nd price percentile after a 626% one-year rally, then fell 32.4% by +1Q with a peak contrarian magnitude of 54.2% — a textbook "mega-upgrade at a price high" case.

7.5 Consecutive upgrades / downgrades (sequence position vs contrarian rate)

Streaks are defined as same-firm / same-instrument / same-direction runs with gaps <=180 days and no intervening opposite revision. The 2nd consecutive call (n=11) shows a 22.2% contrarian rate at +1M and mean peak CM of 12.0%, versus 46.7% and 14.6% for the first call (n=63). Third-or-later calls number 2 (too few to conclude). Repeated revision in one direction is itself a state variable worth tracking: once the market has been re-anchored several times, the marginal information and reversal risk of the next call both change.

7.6 Verifiable consecutive-revision chains

A streak is a run of consecutive same-direction revisions by the same firm on the same instrument, with gaps <=180 days and no opposite-direction revision in between. 序列长度受事件池覆盖限制：媒体提及但缺精确日期或无法核实的调整不在池内，因此序列长度是「可核实序列」的下界。
The 2026-11-07 Yicai story "Goldman cuts its gold target again" is one of these chains: on 11-05 Goldman cut its end-2026 gold target from 4650 to 4150 and end-2027 from 5400 to 4800 (P32), following the 09-18 cut from 5400 to 4650 (P30). The report itself calls it a "second cut in two months". This event was already in the sample, not missed; this table now links it explicitly to P30, and the daily tracker will flag whether each new revision is the Nth consecutive one.

Disclaimer: The above is based on public data and quantitative analysis, is for reference only, and does not constitute investment advice. Markets carry risk; invest with caution. Any investment decision should be made independently in light of your own risk tolerance, financial situation and objectives, consulting a licensed professional where necessary. Past performance does not predict future returns.

Are Goldman Sachs & Morgan Stanley Reverse Indicators?