The analyst ID problem in price target data

Analyst identity in price target data is the join key most datasets get wrong, and it is the quietest way a panel goes bad. In AnaChart’s source records, 7,748 distinct broker-and-analyst identifiers resolve to 7,191 real people. Roughly 557 of those identifiers are the same person counted twice.

If you are scoring analysts, weighting a consensus, or building an experience factor, that gap is not a rounding error. It determines whether the entity you are measuring is a person or an artifact.

The two ways a panel breaks

There are only two obvious keys, and both fail.

Key on the brokerage and you blend everyone who has ever held the seat. A firm that has covered a stock for fifteen years through four analysts produces one continuous series that belongs to nobody. The current analyst may be in year two. The record you are scoring describes thirteen years of work by people who have left.

Key on the raw vendor identifier and the opposite happens. An analyst who moves from one bank to another is issued a new identifier and becomes two people, each with a truncated history. Every tenure or experience factor you compute is then wrong by however long they spent at the previous firm, and both fragments fall below whatever minimum-sample filter you apply.

The fix is a canonical identity: one record per human, the broker retained as a separate attribute so firm effects can still be modelled without being confused for person effects.

Turnover is the reason this matters so much

If analysts stayed put, a firm-level key would be a mild approximation. They do not.

The cohort here is the 2,950 analysts with at least 15 recorded targets who began coverage before 2019, so every member has had time to reach the seven-year mark and most the ten. After twelve months 99.8% are still publishing and after three years 95.6% still are. Then the curve bends: 86.8% at five years, 76.5% at seven, 44.9% at ten.

The steepest stretch, years seven to ten, is exactly where an analyst is senior enough to be promoted out of publishing, recruited to the buy side, or cut in a research budget reduction. Leaving coverage is frequently a promotion, which is why attrition accelerates with seniority rather than at the start.

What it does to a live universe

Zoom to a single heavily covered name and the abstraction becomes concrete.

Distinct analysts who have ever recorded a price target on each name, and how many have recorded none in the twelve months to the April 2026 snapshot.

On Apple, 76% of the analysts who ever covered it are inactive. Any firm-level history of Apple coverage is therefore mostly a composite of people who have moved on. Nvidia is the least turned over of the five at 52%, and the reason is instructive: much of its coverage was initiated relatively recently as the stock became a mandatory holding, so the roster has had less time to churn.

The survivorship trap

Once turnover is this high, panel construction becomes a bias decision, and both defaults are wrong.

Build the panel from currently-active analysts and you drop 39.8% of the population. That subset is not random. It over-represents people who left after a poor run, were cut in a budget reduction, or covered a sector that fell out of favour, which is precisely the tail that makes an accuracy distribution honest.

Build it from everyone but key on the firm and you avoid the survivorship problem by introducing an attribution one, scoring institutions for work done by individuals who are no longer there.

Sample size is the other half of the problem

Turnover does not just move people out of the panel. It caps how much evidence most of them ever accumulate.

Longest spans between first and most recent recorded target, restricted to analysts still publishing within the last twelve months.

These are the records you can score confidently, and they are exceptional. The median analyst in the full population has 26 recorded events across 9 tickers. At that sample a hit rate is not distinguishable from noise, which is an argument for taking the whole population and filtering yourself rather than accepting a vendor’s pre-filtered subset without knowing what it removed.

What to require from any analyst dataset

Person-level canonical identity, stable across employer changes, with the broker as a separate field.

The full population, including analysts who have stopped publishing, so you can construct your own point-in-time universe rather than inheriting someone else’s.

First and last activity dates per analyst, so recency and tenure are computable rather than assumed.

The reference price at the moment each target was issued, without which no accuracy metric can be recomputed or audited.

AnaChart’s dataset carries all four across 833,361 records, 7,191 canonical analysts and 9,686 tickers, from 2004 to 2026 with deep continuous coverage from 2013, delivered via Snowflake, BigQuery or flat file.

Method and limits

Population. Career-span figures use the 4,369 analysts with at least 15 recorded events, which removes lightly covered names whose spans would be meaningless. Span is elapsed time between an analyst’s first and last recorded target.

Survival curve. Restricted to the 2,950 qualifying analysts who began before January 2019 so every member has had the opportunity to reach the measured anniversaries.

Inactivity is inferred from an absence of recorded targets in the twelve months before the April 2026 snapshot. It cannot distinguish retirement from a firm move, a coverage suspension, or a gap in the source record.

Known limit. Career span is bounded by the coverage window, so anyone publishing before 2004, or before dense coverage began in 2013, has their true tenure understated. The ten-year survival figure should be read as a floor.

Frequently asked questions

Why does analyst identity break a price-target panel?

Because the join key most feeds offer is the wrong grain. AnaChart’s source records carry 7,748 distinct broker-and-analyst identifiers that resolve to 7,191 real people, so roughly 557 identifiers are duplicates created when someone changed firm or desk. Key on the raw identifier and one career splits into two shorter ones. Key on the brokerage and four people’s records blend into a single series. Either way the entity you are scoring is not a person.

How fast does the analyst population turn over?

Faster than the coverage universe suggests. Among 4,369 analysts with at least 15 recorded targets, the median career span is 7.3 years, 31.4% span under five years, and 39.8% have published nothing in the last twelve months. Tracking a cohort that began before 2019: 95.6% were still publishing after three years, 86.8% after five, 76.5% after seven and 44.9% after ten.

What does that do to a backtest?

It introduces survivorship bias in whichever direction your panel is built. Restrict to currently-active analysts and you silently drop 39.8% of the population, and the ones you drop are disproportionately those who left after a poor run or were cut in a budget reduction. Include everyone but key on the firm and you inherit the blended record of people who are no longer accountable for it. On Apple, 125 analysts have set a target at some point and 95 of them, 76%, have set none in the last twelve months.

How does AnaChart resolve analyst identity?

Records are tracked against the named individual and consolidated to a canonical analyst name, so a move between brokerages preserves one continuous history rather than creating a second short one. That consolidation is why the dataset reports 7,191 analysts rather than the 7,748 raw identifiers in the source records. The broker field is retained separately, so you can still model firm effects without conflating them with person effects.

Which analysts have the longest continuous records?

John Hodulik spans 22.1 years across 347 targets on 45 tickers. Tristan Gerra, James Ricchiuti and Paul Cheng each span 20.8 years, with Ricchiuti the most prolific at 1,245 targets on 54 tickers. Michael Baker spans 20.7 years across 1,155 targets on 72 tickers. These are the records with enough observations to score confidently, and they are rare: the median analyst in the full population has 26 recorded events across 9 tickers.

What is the minimum sample before an analyst score is meaningful?

AnaChart’s own award rankings require roughly 15 targets before an analyst is eligible, and that is a reasonable floor for external use too. The median analyst never reaches a sample where a hit rate is distinguishable from noise, which is a further argument for carrying the full population and filtering on sample size yourself rather than accepting a pre-filtered feed.

Data as of April 2026. Published 27 July 2026.