gdp_surprise_anchor

1. The number being subtracted is doing most of the work

This is the one that changes how you should read the switch. Excluding the pandemic, the last-quarter figure accounts for 85.1% of the variance of the gap; its correlation with the gap is −0.676, against +0.485 for the current-quarter nowcast. In plain terms: the gap widens more reliably because last quarter was bad than because this quarter is good.

The cleanest way to see it is to compare the two states. On the 957 days the gate was open, the average previous quarter had grown +1.70% and the nowcast read +4.41%. On the 1,552 days it was shut, the previous quarter had grown +3.42% and the nowcast read +2.36%. Both legs move, but the previous quarter moves downward by 1.72 pp into the open state while the nowcast moves upward by 2.05 pp. A weak quarter behind you is close to half the signal.

There is a mechanical reason as well as a statistical one. Quarterly growth is measured against the previous quarter's level, so a quarter dragged down by a one-off — a strike, a weather event, an inventory unwind — mechanically raises the next quarter's growth rate even if underlying demand is flat. The signal cannot tell that rebound apart from a genuine pick-up in activity. "Growth accelerating" and "growth was unusually bad three months ago" produce the same reading.

2. Two legs, and neither stored one is what anyone could have seen at the time

gdpnow_anchor has one leg with a lookahead problem: FRED dates each nowcast by the quarter's start and revises it in place, so our table holds the final answer filed under a date thirteen weeks earlier. This signal inherits that leg — and adds a second one with the same disease. finance_quotes keeps only the latest value of A191RL1Q225SBEA, and BEA revises that series repeatedly: advance, second, third, then annual updates that rewrite years at a time.

Measured against FRED's vintage archives, on 94.0% of the 2,792 evaluable days the published figure our table holds is not the one that was actually on the wire that day. The revision averages 0.83 pp and reaches 4.90 pp. Combined with the nowcast leg's own revision (mean 1.64 pp), a replay of our two tables lands a median 1.22 pp away from the live gap, p90 4.17 pp, worst case 21.88 pp.

As a switch, at the shipped threshold, the two disagree on 25.4% of days — and the error leans the opposite way from gdpnow_anchor's. There, the naive replay traded on days the real signal blocked. Here it is mostly the reverse: 19.9% false-shut against 5.5% false-open. That is the less dangerous direction — a replay that under-trades flatters you less than one that over-trades — but it still means a quarter of the days in any study you run on this signal are the wrong days. The second leg does not cancel the first; it makes the combined error larger than the nowcast leg's alone.

3. A backtest containing this signal is not testing it

The string gdp appears zero times in quants/backtest.py — not in the supported list, not even in the commented _UNSUPPORTED_ENTRY block that documents why other anchors are excluded. Only the catch-all k not in _SUPPORTED_ENTRY stops it, so the backtest runs without the gate and hands you an equity curve anyway. You can tune a strategy that was never once tested with the switch you are about to ship it with.

In fairness it is not silent about it: the run flags unsupported_rules and the UI prints "advanced rule(s) not yet replayed". So it is a lie by omission rather than a lie — but the curve still looks like a result. This is the same defect documented on gdpnow_anchor; both GDP signals are missing from that file together.

4. The threshold is smaller than the measurement error of its own inputs

The gate asks whether a difference exceeds 1.0 pp. Consider what 1.0 pp is worth against each leg.

• The nowcast leg: GDPNow's final estimate of a quarter — its most accurate — misses BEA's first published figure by a mean absolute 0.869 pp over the 60 quarters measured on the gdpnow_anchor page. Earlier it is far worse: the gap to the quarter's final nowcast averages 2.43 pp in the first fortnight of nowcasting against 0.50 pp in the last stretch.
• The published leg: BEA's own advance estimate moves a mean 0.492 pp by the third estimate and 0.883 pp by the latest vintage. It exceeds the entire 1.0 pp gate threshold in 28.6% of the 42 quarters since 2016 — 2025 Q4 alone went from +1.4% to +0.5%.

So a threshold of 1.0 pp is roughly one measurement error of either leg. It is not an absurd number — you have to put it somewhere, and the duty cycle it produces is reasonable — but it should not be read as separating economic regimes. A gap of 1.1 and a gap of 0.9 are not distinguishable states of the world.

5. The docstring promises both feeds fail closed. Only one of them does.

The module says, in capitals, "FAILS CLOSED if either feed is missing/stale". The nowcast leg genuinely does: it parses the timestamp into a real Python datetime and rejects a row whose quarter is over 100 days old or whose poller has not run in 3 days.

The published-figure leg does not. Its check is fetched_at >= datetime('now', '-1800 seconds'), which compares our Python ISO timestamps — 2026-08-12T08:25:01+00:00, with a T — to SQLite's space-separated 2026-08-12 08:00:58. SQLite compares them as text, and "T" sorts after " ", so any row bearing today's UTC date passes. Verified live on production at 08:30 UTC: a row stamped 00:00:02 passes the supposed 30-minute test, and 101 rows were inside the "30-minute window". The real window is "since midnight UTC" — 30 minutes at 00:30, nearly 24 hours at 23:30.

The window was earning its keep, too: the GDPGROWTH feed has five gaps over 30 minutes, the longest 85 minutes, and FRED has failed on this specific series 9 times. It does still fail closed after the UTC date rolls over, so a truly dead feed is caught within a day. This is the same defect already documented on finance_anchor, tropical_anchor and funding_anchor. Not fixed here — this loop does not edit anchors or pollers — but Dave has it.

6. It is called a surprise, and it is not one

In macroeconomics a "surprise" is a published number minus what forecasters expected — the news content of a release, which is what markets actually reprice on. Every surprise index you have heard of is built that way. This signal holds no forecast and no consensus of any kind. It computes a nowcast minus a different quarter's realised figure, which is an acceleration, not a surprise.

Credit where due: the module docstring says exactly that — it calls it "the change in momentum" — and the builder's own description spells out the arithmetic honestly ("the live Atlanta Fed GDPNow nowcast minus the last official GDP print"). The defect is the label sitting on top of an accurate description. A user who reads the field name and skips the paragraph will think they have switched on "trade when GDP beats expectations". They have switched on "trade when this quarter's model output exceeds last quarter's published number by a percentage point" — which can be true in an economy nobody is surprised by, and false in one that has just shocked everybody.

7. The clock in the quarter decides more than you would expect

The gate is open 48.0% of days 30–44 into a quarter and only 25.4% of days 90 and beyond — roughly twice as likely to let you trade early in a quarter as late in one. The average gap follows the same arc: +0.89 pp on days 30–44, drifting down to −0.68 pp on days 90 and later. In other words GDPNow's early readings tend to sit above where the quarter eventually lands, so the gate is at its most permissive when the model has the least data. Why the early readings run high is not something we can settle from our own records, and this page is not going to guess — but the pattern is there in all 2,509 non-pandemic days, and nothing in the product mentions it.

On top of that, the gate is completely dark on 25.35% of all calendar days — 948 of the 3,740 modelled — because the previous quarter's row ages past the 100-day limit around day 8 of a new quarter while the new quarter's first nowcast does not land until about day 30. That behaviour is correct: refusing to treat a 92-day-old final as "now" is exactly right. The defect is that it is silent — no field note, no warning when you save, no log line. Four whole months in our window (Feb 2019, Nov 2025, Feb 2026, Mar 2026) had not one single day on which this gate could be evaluated.

8. What the registry says, and what is actually true

quants/signal_registry --json resolves this signal to "poller": "gdpnow_poller" and three tables: finance_quotes, macro_nowcast, and gdp_surprise_anchor. Two corrections.

• There is no table called gdp_surprise_anchor. The registry harvests backticked words from the docstring, and the docstring backticks the signal's own name in the sentence "bots that explicitly set gdp_surprise_anchor". A phantom, not a fabrication — but it is on the list.
• Naming one poller understates the dependency. This gate needs gdpnow_poller and finance_poller, on completely different schedules. If finance_poller stops, this signal fails closed and every bot carrying it goes silent — and nothing in the registry, the health check or the builder would point you at the right feed.

This is a different failure mode from the predictive.db regex artefact that hit five other signals: here the registry simply has no concept of a signal with two upstreams. The rest of what it reports — cron, replayable false, live_bots 0 — is correct, and here 0 really does mean 0.

9. Configurations that validate but cannot mean what they say

• validate_config bounds thresholds to ±20 pp, so {surprise_below: -30} is rejected — although the gap we reconstructed ranged from −48.8 to +66.7 in 2020. A user trying to write a rule about the only period in the record where this signal moved dramatically is told the number is out of range. The bound was copied from the level gate, where ±20 is roomy; a difference of two growth rates has about twice the range.
• {surprise_above: -5} validates and to_english renders it "GDPNow is running > -5pp ABOVE the last actual GDP print (growth accelerating)". It is a double negative describing a decelerating economy, labelled accelerating, and it is open almost always.
• {surprise_above: 3, surprise_below: -1} validates, is unsatisfiable, and reads back as "accelerating and cooling". The builder at least prints both clauses, which is better than funding_anchor, where the second one is silently hidden. This same validator gap now appears on six pages of this dictionary.