How The Gold Barometer works
The Barometer is one number, from 0 to 100. It comes from seven conditions we measure every day. Each condition gets a score from 0 to 100 by comparing today with every day since 1971. We then average the seven with published weights. This page is the full recipe. Version v1.0.
A barometer does not tell you if it will rain. It tells you the pressure now, and what usually followed pressure like this. The Gold Barometer does the same for gold buying conditions. It reads seven parts, ranks each one against its own past, and averages the ranks. It is a measurement. It is not advice.
1. The seven parts, their weights, and how strong the evidence is
The weights are fixed for v1.0. Retail premiums are new. We only started collecting them in August 2026. That part's weight starts at zero and grows a little every day. It carries nothing until it has 90 days of its own history. It reaches 5% at one year and 10% at two. It never jumps on any single day. The other parts share out the missing weight meanwhile.
The right-hand column shows how sure we are that each part predicts long-term returns. Strong means many decades of evidence. Weak means the part is useful as context, not as a forecast.
| # | Part | Weight | Strength of evidence |
|---|---|---|---|
| 1 | Cash after inflation | 25% | Strong, with one big exception |
| 2 | Entry price | 20% | Strong on both pieces, at different time horizons |
| 3 | Central-bank buying | 15% | Strong for central banks, weak for ETF flows |
| 4 | US dollar | 10% | Moderate |
| 5 | Trader bets | 10% | Weak to moderate |
| 6 | Price jumpiness | 10% | Weak as a predictor of returns |
| 7 | Retail premiums | 0% today (phased in gradually) | Strong on what you pay, weak as a predictor |
2. Rank against its own past
Each of the seven parts gets a rank from 0 to 100. We compare today with every recorded day of that part. The score starts in 1971. Two parts' records reach further back, to 1962 for cash after inflation and 1960 for the entry price. The rank uses all of it. Younger sources start where their data starts. A rank of 80 means today is higher than 80% of all past days. 100 is the highest ever. 0 is the lowest ever.
Ranking makes very different measurements comparable. It also does not blow up when a value spikes. Some sources allow us to publish the rank but not the raw number they own. Ranking respects that rule too. Details are on the provenance page.
Each rank is then flipped so that a higher score means better historical buying conditions. Cash after inflation: low is better for a buyer. Entry price: cheap after inflation and below its long-run path is better. Central-bank buying: rising is better. US dollar: weak is better. Trader bets: few bets left on a rising price is better (buyers do best when others have given up). Price jumpiness: low is better as an entry-quality gauge. Retail premiums: low is better (lower all-in cost).
Every flip is set by the plain economic reason above. None of them is fitted to past gold returns. That matters: the score cannot be shaped to look good after the fact.
3. How the seven turn into one number
The score is an average of the seven parts, but not an equal one. Each part counts in proportion to its weight. Each daily input is smoothed: we take the middle value of its last five days. One odd day cannot swing it. Nothing else. Each part has its own page with the full detail.
A note on the retail premiums part. Our basket only holds dealers whose terms allow price collection. It only holds dealers whose price pages our collector can open without an account. The collector announces itself by name to every site it visits. That excludes most of the biggest US dealers by sales. So the movement of the median from day to day is the signal, not the absolute level. Full details on the provenance page.
4. The five zones
We split the 0 to 100 line into five zones. The words are how we speak about the reading.
| Zone | Band |
|---|---|
| Historically very favorable | 80 to 100 |
| Favorable | 60 to 79 |
| Mixed | 40 to 59 |
| Unfavorable | 20 to 39 |
| Historically very unfavorable | 0 to 19 |
The words say "conditions". They do not say buy or sell or should. We send an email only when the reading crosses into a new zone. The signup form is on the home page.
5. What happens when a source falls silent
Every part has a limit on how old its data can be. Inside that limit, we carry forward the last known value and flag it as ageing. Past that limit, the part drops out for the day. The other six share out its weight so the score stays on the 0 to 100 scale. The home page and provenance page both show "N of 7 parts used today" so anyone can check.
- Daily sources: 7 business days.
- Trader bets, published weekly: 10 days.
- The dollar index, weekly but slower to arrive: 14 days.
- Monthly sources: 45 days.
6. Versioning: no silent revisions
v1.0 is the version we launched with. If we ever change a weight, a flip, or the recipe, the version number goes up. The change goes in the log below. We also run the old and new versions in parallel for 90 days before the new one takes over.
If a published number is ever wrong, the fix goes on the corrections page with the date and the reason. Nothing gets silently rewritten.
Changelog
| Version | Date | Change |
|---|---|---|
| v1.0 | 2026-08-05 | Launch. Seven parts, weights 25/20/15/10/10/10/10. Retail premiums enter on a ramp as their own history grows. |
| v1.0 | 2026-08-12 | The premiums ramp became continuous (no weight before 90 days, then daily growth to 5% at one year and 10% at two). Decided and published before the part carries any weight, so no published reading is affected. |
| v1.0 | 2026-08-11 | Added section 8: the 90% ranges, episode counts and independent-window counts behind every band figure. The same day, separately, three engine faults were corrected (see the corrections log): that correction moved readings, this section addition did not. |
7. Where this framework has been wrong
Every claim below is dated and sourced. The point is to state the limits of the instrument in numbers, not adjectives. If a fact is uncomfortable, it stays.
7.1 The cash-after-inflation link broke down in 2022
The cash-after-inflation part usually moves in the opposite direction to gold. That link was strong for two decades. It stopped working in 2022. The Chicago Fed Letter 464 (2021) already noted that the link was essentially absent before 2001.
RBC Wealth Management reports how much of gold's move the link explained. It was 69% for 1997 to 2004. It was 84% for 2005 to 2021. It fell to 3% for 2022 to 2023. It was 7% for 2024.
JP Morgan Asset Management reports the same slide: about 85% (1990 to 2021) to about 16% since 2022. The link now points the opposite way.
This part is stitched from two sources, and the stitch shows. Before December 2005 we use a calculated stand-in. From then on we use market TIPS. The two disagree by about half a percentage point. At the switch, that gap alone moved this part's score by about 18 points on its 0 to 100 scale. A smoother join is a candidate for v1.1.
Our own test on 1971 to today, monthly, measures the link directly. We measure it between the real rate itself and gold's next 12 months, on a scale from -1 to +1. Minus 1 means the two always moved opposite ways. The claim expects a negative link: low rates, better gold ahead. The link was +0.237 for 1971 to 2000, so the claim failed that whole era. It was -0.401 for 2001 to 2021, then -0.851 for 2022 to 2024. The direction the claim expects has only shown up after 2000.
7.2 The top zone has never happened
Across 667 monthly readings from 1971 to today, the score has never crossed 80. The top zone (80 to 100 "Historically very favorable") has zero observations. We do not publish forward numbers for it. We will only publish them once at least one month lands there and we can see what happened next.
7.3 At 1 year, the zones do not line up
Over one year, medians should rise from a lower zone to a higher one. They do not. The 1-year medians before inflation are: Favorable 5.8%, Mixed 6.5%, Unfavorable 3.7%.
Read as a 1-year buy signal, the instrument fails. The short-run reason is simple. When gold has just fallen a lot, the score drops with it. When gold has just risen a lot, the score rises. In the typical month, short moves and the score point opposite ways.
Over five years, the long-run thesis does hold. Median 5-year real returns are 48.5% in the Favorable zone and -15.8% in the Unfavorable zone.
7.4 The two bottom zones sit out of order, and both lost ground
Over five years after inflation, the Unfavorable zone (20 to 39, 152 months) shows -15.8%. The zone below it (0 to 19, 27 months) shows -7.7%. The order is wrong, and both are losses.
The bottom zone's months bunch into a few short episodes. Laid end to end, they do not even fill one five-year stretch. Its median is a description, not a signal. Section 8 shows why no honest range can be put around it.
An earlier version of this section reported the bottom zone at +22.1% after inflation. That figure came from a coding error at the December 2005 source switch described above. It was fixed on August 11, 2026. The corrections log carries the full account.
7.5 The trader-bets bet-against-the-crowd rule is practitioner lore
The published bet-against-the-crowd rule watches mining firms and big metal dealers as they insure against a price fall (Wang, 2003). Our part uses the professional funds' bets because the data is clean and goes deeper. That is the reason its weight is only 10%. The caveat is public.
7.6 Price jumpiness and premium levels do not predict returns
There is no published evidence that levels of price jumpiness predict gold's next return. There is no such evidence for retail premium levels either. Both parts are risk and cost gauges. Neither is a forecast.
7.7 The retail premiums part is US-only
Our retail premiums basket is US-only. In March 2020, Western coin demand rose 36% against the same months a year earlier (World Gold Council, early 2020). At the same time, gold in India sold about 70 dollars an ounce below the world price (Singapore Bullion Market Association). One global premium number does not exist. This site publishes the US read.
7.8 Fund flows follow the price, they do not lead it
ETF flows move at the same time as the price rather than ahead of it (World Gold Council flow data). Inside the central-bank buying part, the ETF signal and central-bank purchases carry equal halves. The central-bank half has a documented cause: central banks bought more gold after watching sanctions freeze other countries' money. The ETF half is the weak half, and we say so.
8. How sure can we be?
Every middle result above comes from stretches of history that overlap. The honest question is how much those results would move if the record were replayed. So we replay it.
The method: draw the 55-year record again 10,000 times, in blocks as long as the horizon, so that runs stay runs. Recompute every zone's middle result in each replay. The middle 90% of those 10,000 answers is the range we publish.
| Zone | Months | Separate episodes | How far ahead | Middle result | Where 9 of 10 landed | Non-overlapping stretches |
|---|---|---|---|---|---|---|
| Favorable | 110 | 27 | 1y nominal | 5.8% | -4.2% to 13.7% | 9.17 |
| Favorable | 110 | 27 | 5y nominal | 71.3% | 38.8% to 136.6% | 1.83 |
| Favorable | 110 | 27 | 5y real | 48.5% | 23.5% to 102.6% | 1.83 |
| Mixed | 379 | 43 | 1y nominal | 6.5% | 0.5% to 15.5% | 31.33 |
| Mixed | 379 | 43 | 5y nominal | 43.8% | 4.3% to 80.2% | 5.75 |
| Mixed | 379 | 43 | 5y real | 19.4% | -10.8% to 40.8% | 5.72 |
| Unfavorable | 152 | 21 | 1y nominal | 3.7% | -2.9% to 16% | 11.92 |
| Unfavorable | 152 | 21 | 5y nominal | -0.2% | -11.3% to 8% | 2.1 |
| Unfavorable | 152 | 21 | 5y real | -15.8% | -27.7% to -9.2% | 2.1 |
| Historically very unfavorable | 27 | 4 | 1y nominal | 9.7% | -12% to 58.5% | 2.25 |
| Historically very unfavorable | 27 | 4 | 5y nominal | 9.2% | too few months to say | 0.45 |
| Historically very unfavorable | 27 | 4 | 5y real | -7.7% | too few months to say | 0.45 |
Too few months to say means the zone holds fewer months than one horizon. Under one independent observation, no range is honest.
Two admissions come out of this table. 3 of the four one-year ranges hold both a loss and a gain. The one-year rows carry no usable signal, as section 7.3 already said. And the bottom zone cannot be ranged at five years at all, because its months, laid end to end, do not fill one five-year stretch.
What survives all of that is the claim the site rests on. In 100% of the 10,000 replays, the Favorable zone beat the Unfavorable zone at five years after inflation. The gap between them was 64.3 percentage points of five-year return after inflation. In 9 of 10 replays it sat between 36.9 and 119.7.
The replay code and the full monthly series sit in the public repository. The reshuffling starts from a fixed number, so a rerun produces the same 10,000 histories. Anyone can reproduce this table to the digit.
9. What we looked at and left out
The v1.0 set is deliberately small: seven parts, each with a plain reason and a public evidence grade. The candidates below were looked at and left out. Some are marked as candidates for v1.1.
| Candidate | Status | Reason |
|---|---|---|
| What markets expect inflation to be over 10 years (10-year breakevens) | Rejected | Already inside cash after inflation. Adding it would count the same thing twice. |
| M2 money-supply growth (how much money is being printed) / government overspending (deficits), which can cheapen money | Rejected | Already inside the entry-price part, which uses inflation-adjusted valuation. Adding money-supply data would count the same thing twice. |
| The extra yield long bonds pay over short ones (the 10-year term premium) | v1.1 candidate | Overlaps with the cash-after-inflation part. Logged for v1.1 as a possible overlay. |
| Geopolitical risk index (Caldara-Iacoviello) | v1.1 candidate | A known driver of gold at event dates. Hard to time day-by-day. Logged for v1.1 as an event overlay, not a full part. |
| Miner all-in sustaining costs (AISC) | Rejected | Quarterly, laggy, and industry-reported. Weak evidence that it predicts returns. |
| Seasonality (month, day-of-month) | Rejected as a part | A real effect, but tiny. Any weight big enough to notice would overstate it. Covered in our seasonality article instead. |
| Wholesale market stress gauges (the cost to borrow gold, and the paper-to-physical price gap) | Rejected | Data is licensed or fragile. The retail premiums part already stands in for physical tightness at the consumer end. |
10. Sources and credit
We take each number from the office that first publishes it. Our six sources are:
- Board of Governors of the Federal Reserve System
- US Bureau of Labor Statistics
- US Commodity Futures Trading Commission
- International Monetary Fund
- World Bank
- Cboe
The provenance page lists each source, its licence, and its freshness limit. Its licence lets us use the Cboe Gold Volatility Index in our sums, but not republish its numbers. We do not republish its raw values.
See the history page for the full 1971 test. It has per-zone what-happened-next results, worst drops along the way, the buy-a-little-each-month comparison, and the failures in full.