I. Why standard tools fail
Two structural shocks in 2022 broke assumptions that had held for over a decade. The tools most investors rely on were designed for a world that no longer exists.
Trigger 1: The end of the zero-rate era
For over a decade, the zero-rate environment was the invisible constant behind almost every portfolio decision. Bonds yielded little, but they rose reliably when equities fell — the 60/40 promise worked. The Fed's rate pivot in 2022 destroyed that assumption within months.
TLT lost more than 30% in the same year the S&P 500 fell more than 18%. The 60/40 portfolio suffered its worst drawdown in decades — not because the idea was wrong, but because the structural precondition (inverse correlation between equities and bonds) had ceased to exist. Most investors noticed only after the damage was done.
Trigger 2: The emergence of AI
ChatGPT (November 2022) and the AI wave that followed reshuffled the entire technology sector. Previous winners — SaaS, fintech, cloud-native pure plays — became losers. New winners — semiconductors, infrastructure, hyperscalers — exploded. The "Magnificent 7" pulled the S&P 500 higher, but the equal-weighted version of the index told a different story.
Index returns and individual stock returns diverged as sharply as they had in years. Anyone looking only at the broad index missed that the market was undergoing a massive structural transformation beneath the surface.
The problem with the standard toolkit
What does the typical experienced investor have available? RSI, MACD, Stochastic, Bollinger Bands, ADX, CCI, Williams %R, OBV — an arsenal of technical indicators available on every TradingView, Barchart, or broker platform.
The problem: most of them say the same thing in different ways.
RSI and Stochastic both measure momentum and overbought/oversold conditions. MACD and moving-average crossovers both measure trend strength. Bollinger Bands and ATR both measure volatility. CCI and Williams %R are, at their core, variants of the same concept. Twenty tools, but really only three or four dimensions: trend, momentum, volatility, volume.
And they all share three fundamental limitations:
Bloomberg Terminal can do all of this — at $25,000 per year, designed for institutional trading desks, not for experienced retail investors or boutique advisors.
Not "is QQQ overbought?" but "WHY is QQQ overbought — because the entire market is running hot (systemic), or because tech is specifically decoupling from the rest (divergence)?" The answer determines whether to sell, hold, or even add — and standard tools cannot even ask this question.
- RSI, MACD, Stochastic on QQQ
- All say "overbought" — same signal, three times
- No view of GLD, TLT, BTC simultaneously
- No relationship tracking between pairs
- Backward-looking: shows what happened
- 5 orthogonal dimensions per instrument
- Cross-asset pair spreads reveal divergences
- Cointegration tracks if relationships are alive
- Cascade Tracker watches the domino sequence
- Forward-looking: conditions building toward the next move
II. The architecture: two pillars, three levels
Money Temperature combines a proprietary scoring model with institutional-grade analytical methods — made accessible through daily automated dashboards with narrative commentary.
Pillar 1: The Temperature Model
Developed from first principles analyzing gold price behavior and participant stratification, the Temperature Model scores each instrument on a 0–100 scale across five dimensions. The core thesis:
The model maps market participants to moving averages by their "pain threshold": the 20-day MA represents hot money (day traders, options), the 50-day represents trend followers (CTAs, momentum ETFs), and the 200-day represents the strategic floor (central banks, sovereign wealth funds, pensions). The temperature reading quantifies where in this stack price currently sits, how fast it's moving, and how fragile the structure is.
Five dimensions, one composite score
A reading above 80 means an instrument is "boiling" — all dimensions at extremes. Between 50 and 79 is "warm" — a healthy trend with support. Between 20 and 49 is "cool" — weakening momentum. Below 20 is "freezing" — sell-off or base formation.
The scoring is relative, not predictive. A reading of 80 on GLD means gold is running hot across all five dimensions — not that it will reverse. The signal's value lies in cross-reading: when SPY and GLD both read above 70 while TLT stays below 30, you are looking at a specific regime configuration.
Eight instruments, four strategic pairs
| Pair | Instruments | What it reveals |
|---|---|---|
| Risk vs. Haven | SPY − GLD | Risk appetite versus flight to safety |
| Tech Premium | QQQ − SPY | Speculative excess in the technology sector |
| De-dollarization | BTC − GLD | Digital versus physical inflation hedge |
| Duration Rotation | TLT − SPY | Bond rotation versus equities |
The temperature spread between paired instruments — not the individual reading — reveals the regime dynamic. An extreme spread (beyond ±25 points) signals structural stress.
Pillar 2: Institutional methods, made accessible
Six analytical methods drawn from academic research and institutional quant practice, running daily on the same instrument basket. Each exists in textbooks and on quant desks — but nobody offers them as a daily, narrated dashboard for the experienced individual investor or boutique advisory firm.
| Method | What it measures | Origin |
|---|---|---|
| Engle-Granger | Do two time series share a stable long-run relationship? | Nobel Prize 2003 |
| Spread Half-Life | How quickly does a stretched pair revert to equilibrium? | OLS on lagged spread |
| Hurst Exponent | Mean-reverting, random walk, or trending? | Hurst 1951 (R/S method) |
| Variance Ratio | Confirmation: does variance scale sub-linearly or linearly? | Lo & MacKinlay 1988 |
| Absorption Ratio | Systemic fragility — are markets moving in lockstep? | Kritzman 2011 (MIT Sloan) |
| Spread Z-Score | How far is the current relationship from its normal state? | Standard statistics |
Three levels of application
The identical analytical pipeline applies across three levels. Only the ticker universe changes.
| Level | Instruments | Question |
|---|---|---|
| Macro | SPY, QQQ, TLT, UUP, GLD, BTC, VEU, EEM | "What is happening in the big picture?" |
| Sector / Regional | 11 SPDR sector ETFs, 15+ country ETFs | "Where is the money rotating to?" |
| Thematic | IGV, SOXX, ARKK + Closelook indices | "Is our thesis confirmed by the market?" |
III. Reading the methods: academic vs. investor
Every method in the toolkit is presented through three systematic perspective shifts — from how textbooks describe them to how investors can actually use them.
Flip 1: From snapshot to film
Textbooks treat each method as a single test at a single point in time. The Closelook approach tracks the value over time. Not the Hurst exponent — but the Hurst drift. Not whether cointegration holds — but whether the p-value is rising or falling. Not the Absorption Ratio level — but its rate of change. The gradient is always more informative than the level.
Flip 2: From abstract to named
PCA outputs "Factor 1, Factor 2." Closelook names them: central banks, hot money, macro hedge, inflation, jewelry. Named factors connect to real-world events. "PC1 loading increases" becomes "central banks are buying more aggressively since the China policy decision."
Flip 3: From "is" to "will be"
Textbook methods are descriptive: "cointegration holds." Closelook adds the predictive layer through combination: "Cointegration holds for now, but the half-life is rising and Hurst is turning — expect the break within two to four weeks." No single indicator is predictive. The combination of several signaling the same thing IS the forecast.
Method by method
Cointegration → "Is my assumption still valid?"
The textbook tests once and says yes or no. Closelook tracks the p-value over time as a rolling series. A p-value rising from 0.01 to 0.07 over six weeks shows a dying relationship BEFORE it formally breaks.
Hurst Exponent → "Will it snap back or keep running?"
A single Hurst value is static. Closelook shows the Hurst drift: was it 0.4 last month, 0.45 the month before, now 0.72? That is not noise — that is a regime change in spread character.
Absorption Ratio → "Can I trust my portfolio today?"
The AR change matters more than the level. AR typically rises two to four weeks BEFORE major drawdowns (Kritzman 2011) — not because correlations predict crashes, but because institutional traders unwind their hedges first (raising correlations) and THEN sell their positions (moving prices). Rising correlation IS the hedge unwind. The sell-off comes after.
Granger Causality → "Who is in charge — and since when?"
Most of the time, the Granger direction is stable: SPY leads, GLD follows. When the direction reverses — gold suddenly leading — it is a rare event with high signal value. It means the market is pricing in something not yet visible in equity prices. Closelook tracks not the direction, but the direction change.
Tail Dependence → "Does my hedge work when I need it most?"
Pearson correlation is an average. The average lies. Normal days: SPY and GLD correlate at −0.1 (mild diversification). But on the worst 5% of days: correlation jumps to +0.4. The hedge fails precisely when it is needed most. This is the central insight every 60/40 investor needs after 2022 — visible only through tail analysis, never through standard correlation.
Factor Attribution → "Who is driving the price — and how is that changing?"
The double flip: abstract factors (PC1, PC2) become named buyer groups. And static becomes dynamic — not "who drives the price" but "how is who drives the price changing?" The gradient of factor shares is the real product.
IV. Case study: Gold factor attribution
A worked example demonstrating all three perspective flips — using gold as the prototype for dynamic buyer-group tracking.
Five named buyer groups
Instead of abstract principal components, we identify the actual participant groups whose behavior drives the gold price:
| Group | Character | Timing |
|---|---|---|
| Central Banks | Inelastic, strategic, price-insensitive | Lagging (months) |
| Hot Money | Volatile, elastic, momentum-driven | Leading (2–3 weeks) |
| Macro / Crisis | Event-driven, binary | Coincident |
| Inflation | Slow trend, stable | Stable (low variance) |
| Jewelry / Industrial | Seasonal, price-sensitive | Cyclical (disappears at high prices) |
Regime phases
Use the time slider to move through the regime transition. Watch how central bank dominance fades as hot money takes over — and the warning signals that emerge.
Buyer group timing: lead, lag, coincident
Not all buyer groups move at the same speed. Understanding who moves first is the predictive edge — if you can see hot money accelerating while central banks decelerate, the regime is shifting before price confirms it.
The same framework applies to semiconductors (CapEx cycle vs. speculation), Bitcoin (institutional vs. retail vs. ETF flows), and emerging markets (dollar vs. carry trade vs. commodity demand).
V. Use cases: 13 investor questions across three levels
Each question is answered not by a single method but by a stack of three to five methods that confirm or contradict each other. Agreement means high conviction. Disagreement is equally valuable — the honest answer is "unclear."
| # | Investor question | Level | Temp. | Spreads | Engle-Gr. | Half-life | Hurst | Var. ratio | Absorpt. | Z-score | Cascade | Granger | Correl. | GARCH |
|---|
Level 1: Macro regime
Cointegration SPY/TLT shows whether the equity-bond relationship is alive. Half-life shows how quickly deviations correct. Hurst confirms: mean-reverting (relationship intact) or trending (relationship dying). Rolling correlation shows whether they move inversely again or fall together. Tail dependence answers the core question: do bonds protect in a crash — or crash with equities?
QQQ-SPY spread negative plus SPY-GLD spread negative means tech is cooling while gold heats up. Granger causality shows whether QQQ weakness precedes gold strength (causal direction) or whether both respond to the same external factor. Lead-lag cross-correlation quantifies: "Gold typically reacts 3–5 days after QQQ weakness."
Absorption ratio above 60% means instruments are moving in lockstep — one shock propagates everywhere. The fragility dimension shows which individual instruments sit on "dry volume plus Bollinger squeeze." The correlation heatmap reveals whether the high AR is driven by one pair or is broad-based. GARCH shows: will the low volatility persist or normalize?
Cascade tracker: which dominoes have already fallen? The sequence is the signal — crypto first, tech next, broad equity after. Rising half-lives across multiple pairs means relationships are losing elasticity. Hurst drifting from below 0.5 to above 0.5 means spreads are becoming trending rather than reverting. Markov gives a probability: "73% bear regime." A Granger causality reversal means "gold stopped following equities and is now leading — that occurred before the last three regime changes."
GLD and BTC both hot plus UUP cold signals an active de-dollarization bid. Cointegration GLD/UUP breaking means the gold-dollar inverse is no longer reliable — a structural break. CFTC data shows whether commercials (central banks) are buying gold or whether it is purely speculative.
Level 2: Sector and regional rotation
Level 3: Thematic / Closelook indices
VI. Expanding the toolkit
All methods currently live run on OHLCV data from a single source at €20 per month. The expansion path requires zero additional data spend for the first two tiers.
| Method | Investor question | Status |
|---|---|---|
| 5-Dimension Temperature | How hot is this instrument? | Live |
| Pair Spreads + Regime | What regime are we in? | Live |
| Engle-Granger | Is this relationship still alive? | Live |
| Half-Life + Hurst + VR | Will it snap back or keep running? | Live |
| Absorption Ratio | Is the market fragile? | Live |
| Cascade Tracker | How far has the domino effect gone? | Live |
| Granger Causality | Who is leading whom? | Tier 1 |
| Rolling Correlation | Is my diversification working? | Tier 1 |
| Lead-Lag Analysis | Who moves first, by how many days? | Tier 1 |
| Johansen (multivariate) | Does the system work as a whole? | Tier 1 |
| Markov Regime-Switching | What's the bear market probability now? | Tier 2 |
| GARCH | Is the calm before the storm? | Tier 2 |
| Kalman Filter | How is the relationship evolving in real time? | Tier 2 |
| Tail Dependence | Does my hedge work in a crash? | Tier 2 |
| CFTC COT Integration | What are the institutions doing? | Tier 3 |
| VIX Term Structure | Is the market pricing fear correctly? | Tier 3 |
| Factor Attribution (PCA) | Who is driving the price? | Tier 3 |
The data pipeline
The entire system runs on Cloudflare's edge infrastructure. A single Python process pulls OHLCV data daily, computes all scores and cointegration tests, and writes three JSON files to R2 storage. Lightweight proxy workers serve the data to the frontend with CORS and caching. No database, no server, no moving parts beyond a cron job.
Explore the live dashboards
Money Temperature, Cointegration Monitor, and the Agentic Winners 25 dashboard — updated daily.
Open Lab →Money Temperature is a research signal, not investment advice. Past cointegration states, temperature readings, and factor attributions do not predict future outcomes. Closelook Venture GmbH publishes research and maintains reference portfolios. Terms · Privacy