← Daybreak
GOLD × THE S&P 500, HONESTLY · A DATA-DRIVEN OVERVIEW

Gold and stocks: friend, hedge, or stranger?

0 years·0 models

1 question: friend, hedge, or stranger?

Gold and stocks have no fixed relationship. Over the years their together-ness has swung from clearly opposite to clearly in-step and back. We map when they couple, when gold actually protects, why — and where the two sit right now.

Built from 30 models on 55 years of daily history (1971 → mid-2026, the era since gold floated), with public inflation and interest-rate data merged in. We sell no predictions. The job is to separate what the data supports from the “gold is a hedge” cliché. Observation, not advice.

THE WHOLE PICTURE

Two long climbs that never quite move in step.

Since 1971 both gold and the S&P 500 have climbed enormously. Shown on a log scale — the only honest way to compare percentage moves across decades — they look like two parallel success stories.

But the lower ribbon, their year-to-year together-ness, never sits still. It has swung all the way from about −0.28 (moving opposite, around 1973) to +0.28 (moving in step, around 2020) and back. There is no fixed “gold vs stocks” number — only regimes that come and go.

Note: before 1971 the dollar gold price was fixed by law (1833–1971), so there was no market relationship to measure. The story starts when gold floated.

overview

loading overview
Both lines climb across the whole era, yet the together-ness ribbon below them flips from clearly negative to clearly positive and back. The famous crashes are marked — the relationship behaved differently in each.
PART 1 — THE LINK THAT WON'T SIT STILL

The relationship is not low. It's unstable.

You’ll often hear that gold and stocks are “uncorrelated.” That number — close to zero on average — is true, and badly misleading. It’s the average of a relationship that swings violently between moving together and moving opposite.

Every model in this section measures the same together-ness a different way. They all reach the same verdict: it moves in persistent moods, it drifts, and it occasionally resets hard.

corr_rolling

loading corr_rolling
The plain baseline. Over a one-year window the together-ness has run from −0.28 to +0.28; whichever window you pick, the number never sits still. Currently about +0.04.

corr_dcc

loading corr_dcc
A live gauge that rescales each day by how turbulent it was turns earlier than the blunt one-year average — but tells the same story: the relationship keeps swinging through zero.

corr_hmm_joint

loading corr_hmm_joint
There isn’t one correlation — there are a few persistent moods (decoupled, mild, coupled), each lasting weeks to months. Today sits in the mild mood.

corr_wavelet_coherence

loading corr_wavelet_coherence
Together-ness even depends on the horizon. In some eras the multi-year swings line up tightly while the month-to-month moves don’t; in others it’s the reverse. A map, not a clock.

corr_kalman_beta

loading corr_kalman_beta
Gold’s sensitivity to the S&P isn’t a constant. It has drifted from about −0.20 to +0.14 and crossed zero repeatedly — a mild diversifier in some eras, a fellow-traveller in others.

corr_copula_tv

loading corr_copula_tv
Everyday together-ness and the tendency to have extreme days together are two separate facts. There are stretches where day-to-day they look unrelated yet still lurch together on the wild days.

corr_regime_switch

loading corr_regime_switch
The together-ness flips between two regimes that each last about 73 to 93 days on average. Persistence is real — but the switch is only obvious in hindsight.

corr_changepoints

loading corr_changepoints
The relationship didn’t drift smoothly — it reset at four identifiable moments (near 1980, 1984, 1998, 2007), each starting a multi-year stretch with its own average. Clear only after the fact.
PART 2 — IS GOLD A SAFE HAVEN?

Sometimes. And it protects by stepping aside, not by fighting back.

The hope is that gold zigs when stocks zag. The data says something quieter: gold mostly just does its own thing. On the days stocks fall hardest, gold tends not to fall with them — that’s the cushion. It protects through decoupling, not opposition.

And that cushion is episodic. Gold held up in inflation and rate-shock bears; it was dragged down in the 2020 cash-grab before recovering. A haven, but a conditional one.

haven_event_study

loading haven_event_study
In five of five big stock crashes gold rose while stocks fell — strongest in the 1973–74 inflation bear (gold +144%). But in the 2020 cash-grab it dropped about 9% at its worst before recovering.

haven_exceedance_corr

loading haven_exceedance_corr
On the S&P’s worst 10% of days, gold has on average risen about +0.08% — a genuine cushion when it hurts most. The together-ness stays low in both tails; the cushion, not the correlation, is the haven signal.

haven_tail_dependence

loading haven_tail_dependence
When the S&P has one of its worst days, gold has its own worst day the same day only about 6% of the time — close to the 5% you get from two unrelated assets. They rarely crash together.

haven_asym_beta

loading haven_asym_beta
Gold’s sensitivity to stocks hugs zero in every market state — down days, up days, calm, wild. “Haven only when you need it” shows up as gold simply not following stocks down, not as a dramatic flip.

haven_vol_conditioned

loading haven_vol_conditioned
Sorted from calmest markets to wildest, gold keeps going its own way and its own return mostly holds up. The hedge shows up most when the market is scariest — which is when you need it.

haven_classifier

loading haven_classifier
In the months the S&P fell more than 2%, gold rose to the rescue in 55% of them. What separates rescue from let-down: low real interest rates and hot inflation. A description of conditions, not a tradeable signal.
GOLD AS A HAVEN — THE EPISODIC RECORD
0%Gold's gain through the 1973–74 inflation bear
0%Gold's worst dip in the 2020 liquidity crash (then recovered)
0%Of the S&P's worst days that gold also crashed on
0%Of 2%-down stock months where gold rose to the rescue
PART 3 — CAN EITHER PREDICT THE OTHER?

Almost nothing. And we put that first on purpose.

If gold led stocks (or stocks led gold), you could trade it. So we tested every forecaster the honest way — train on the past, predict the future it hadn’t seen, roll forward, repeat. The benchmark to beat is the dumbest guess there is: “tomorrow looks like today.”

Nothing beats it at a day, a month, or a quarter — the AI models included. The one real edge is subtle, and it isn’t a price call.

backtest_skill_heatmap

loading backtest_skill_heatmap
Mostly slate: across 50+ years of rolling tests, letting a model see the partner asset adds essentially nothing. The only blue cells are long-horizon drift bets that the up-trend continues — not a cross-asset signal.

predict_leadlag

loading predict_leadlag
At every daily lag out to a month, one market’s move today tells you nothing reliable about the other’s move days later. Neither gives early warning of the other.

predict_granger

loading predict_granger
A detectable statistical link exists at most lags — but with 14,000+ days, even a tiny pattern shows up. The practical effect is near zero: knowing last week’s gold move shifts the stock forecast by a fraction of a basis point.

predict_transfer_entropy

loading predict_transfer_entropy
The macro backdrop (the real interest rate) carries more information into both gold and stock moves than either asset carries into the other. The shared driver is the macro state, not a cross-asset signal.

predict_var_vecm

loading predict_var_vecm
There is no long-run price link tying the two together. When gold and stocks drift apart, no mechanism forces them back — they are two independent long-run trends.

predict_gbm_direction

loading predict_gbm_direction
A model calling next month’s stock direction is right 68% of the time — which sounds impressive until you see it’s a base-rate artifact (the market is usually up) that collapses across decades. Not a tradeable edge.

predict_deepnet

loading predict_deepnet
A deep-learning forecaster, then the same model handed gold and macro as extra inputs. The partner data adds essentially no short-horizon edge — the lines barely move.

predict_foundation

loading predict_foundation
A large pre-trained forecasting model, run with and without gold as a partner series. Adding gold changes its skill by about zero. Even the biggest models can’t squeeze a cross-asset price signal out of this pair.

corr_forecast_skill

loading corr_forecast_skill
The one real edge: together-ness itself is partly forecastable. A regime-persistence read beats “next = now” by +5% on error and calls the sign of next month’s relationship 54% of the time. The state persists, the prices don’t.

crisis_signal_panel

loading crisis_signal_panel
Using only data available at the time, the relationship confirmed crises more than it predicted them — co-crash stress and regime flips are clearest once a crisis is already underway.
PART 4 — WHY THE RELATIONSHIP MOVES

There's an engine under it — the real interest rate.

The relationship isn’t random. The clearest force behind it is the real interest rate — the long bond yield minus inflation. When real rates are high, gold and stocks couple; when they go deeply negative, the two decouple. Inflation flips gold’s character on top of that.

But here’s the honest counterweight: this engine explains the slow lean of the relationship, not its month-to-month swings. Strip the macro state out and the leftover link barely changes.

macro_realrate_corr

loading macro_realrate_corr
Across the whole 1971–2026 cloud, together-ness rises about +0.006 for every 1 percentage point the real rate climbs. Deeply negative real rates (the 1970s, post-2008, 2021–22) went with decoupling; high real rates went with coupling.

macro_inflation_regime

loading macro_inflation_regime
Gold is a different asset in each inflation regime. In low and rising inflation it drifts mildly with stocks; once inflation runs hot the link flips negative — gold pulls away from stocks exactly when inflation is the danger.

macro_partial_corr

loading macro_partial_corr
The honest limit: after statistically stripping out the real rate and inflation, the leftover link is essentially unchanged (0.22 vs 0.21). The macro engine tilts the relationship slowly — it does not author the swings.

macro_factor_pca

loading macro_factor_pca
One shared force explains about 46% of the joint movement of gold, stocks, rates and inflation — above the 25% of four unrelated series, but nowhere near a master switch. They share a current, loosely and unequally.

macro_real_overlay

loading macro_real_overlay
In inflation-adjusted terms both assets sit high entering mid-2026: real gold set a fresh all-time high in February 2026, and the stock valuation gauge sits in the 97th percentile of its own history. Two expensive assets, not one cheap and one dear.
PART 5 — WHERE WE ARE NOW · MID-2026

An ordinary mood, in an expensive market.

Putting today through every model: gold and stocks are in their mild mood— their one-year together-ness sits around +0.04, neither a strong hedge nor a strong co-mover. The backdrop is rising inflation with positive real rates.

The unusual part isn’t the relationship — it’s the valuations. Both gold and stocks sit historically rich at the same time. The months that most rhyme with now cluster in late 2004 to early 2006.

state_embedding_analogs

loading state_embedding_analogs
On a map where similar market moods sit close together, the months that most rhyme with mid-2026 are Nov 2004, Oct 2004, Jan 2006 and their neighbours — drawn mainly from the 2000s. An observation about which eras look alike, not a prediction.

state_conditional_dist

loading state_conditional_dist
What usually followed regimes like today’s: over the next year gold returned a median +9% and the S&P +11%, gold up 66% of the time and the S&P 74%. A base-rate range on ~13 independent episodes, not a forecast.

state_current_readout

loading state_current_readout
The dated read-out: today’s mood, the inflation and real-rate backdrop, how rich each asset is versus history, and the historical company this setup keeps. Observation, not advice.
WHAT USUALLY FOLLOWED REGIMES LIKE TODAY'S — NEXT 12 MONTHS
0%Gold's median return (8-in-10 range −12% to +52%)
0%S&P 500 median return (range −13% to +31%)
0%Of those episodes gold was up a year later
0 episodesTruly independent past matches — a thin base, stated plainly
PUTTING IT TOGETHER · MID-2026

The honest map.

This is not advice to buy or sell gold or stocks. It is a map of when the two couple, when gold actually protects, and why — and where the pair sits today. Observation, not advice.

The link

No fixed relationship. Their together-ness swings from clearly opposite to clearly in-step and back, in persistent moods. The 'uncorrelated' cliché is a time-average of an unstable, regime-driven process.

Safe haven

Episodic, and it works by decoupling, not opposing. Gold rarely crashes on the same day as stocks (~6%), and on the worst stock days it tends to hold its ground — strongest in inflation and rate-shock bears, weakest in liquidity panics.

Prediction

Almost none. Nothing beats 'tomorrow looks like today' at short horizons — AI and foundation models included. The one real edge is that the relationship's mood persists, worth about +5% over a naive next-month guess. Not a price call.

Why it moves

The real interest rate is the engine: high real rates couple the two, deeply negative rates decouple them, and hot inflation flips gold's character. But the engine sets the slow lean — it does not author the monthly swings.

Where we are

Mid-2026: a mild mood (together-ness ~+0.04) against rising inflation and positive real rates, with both gold and stocks historically rich. An ordinary-looking relationship in an expensive market.

LIMITATIONS — READ THESE
  • Two price series (gold and the S&P 500) plus public inflation and interest-rate data. No options, breadth, or flow data.
  • The era is 1971 onward — since gold floated. Before that the dollar gold price was fixed by law, so there was no market relationship to measure.
  • The 'what usually followed' base-rate rests on only ~13 truly independent past episodes. It is a record of what happened, not a validated forecast — and it does not say when.
  • The structural-reset and regime dates are found in hindsight. In real time you'd notice a break only months after it began; this dates resets, it does not warn before the next one.
  • The relationship confirms crises more than it predicts them. Its haven role is conditional — it held up in some crises and fell with stocks in others.
  • Descriptive, not predictive. These models describe what happened; they don't tell you what's next.

Descriptive research, not financial advice. Both gold and the S&P 500 can fall hard and stay down for years — as both have more than once in this dataset.

Appendix — every chart & model inventory →