Intermarket analysis in practice — one question, comparable observations
A useful intermarket note starts with a question, not a full screen of indicators. We will examine one hypothetical GBP/USD move, calculate what changed in relative yields and check which figures were actually available when the note was written. The result is an explanation that can be challenged, rather than a score built by counting agreeing charts.

Give the observation a question and a deadline
You are reviewing a fall in GBP/USD after a scheduled announcement. Before adding another chart, state the question: did comparable short-maturity yields move against sterling during the same window? This asks about one proposed transmission channel. It does not assume that the announcement caused the whole price move or that the relationship must hold every time.
Use a hypothetical observation day, with every time in UTC. The announcement is scheduled for 14:10. We choose 14:00 and 14:20 snapshots before looking at the outcome. The spot figures are teaching midpoints, while the two yield series use a declared, comparable two-year government-bond convention. Actual work also needs instrument identifiers, a source, quotation side and a record of revisions.
Calculate the relative movement before naming the mechanism
| Observation time (UTC) | UK 2Y yield | US 2Y yield | UK minus US |
|---|---|---|---|
| 14:00 | 4.10% | 4.35% | −0.25 percentage points |
| 14:20 | 4.16% | 4.44% | −0.28 percentage points |
The UK yield rises by 6 bp and the US yield by 9 bp. The UK-minus-US spread changes by 6−9=−3 bp. It becomes more negative, moving from −25 bp to −28 bp. “UK yields rose” alone would miss that relative movement. Adding the two yield rises would answer neither the spread question nor a defined measure of support for sterling.
Let the corresponding GBP/USD midpoints be 1.2700 and 1.26619. The change is 1.26619/1.2700−1=−0.003, or −0.30%. This movement and the decrease in the signed spread are consistent with our proposed channel. They do not identify it as the cause: another announcement, funding conditions or a separate sterling-specific development may have mattered. A government-bond yield also includes more than the expected policy-rate path.
Observation time and availability time are different fields
Suppose the spot snapshot reaches your file at 14:20:02, but both yield snapshots arrive at 14:21:00. Your note has a decision cutoff of 14:20:30. The 14:20 yield measurements exist in the eventual dataset, but were not yet in your information set at the cutoff. You may use them for a later reconstruction, clearly dated after their arrival. Using them in the earlier note would introduce information from the future.
This distinction also matters with public daily series. The ECB explains that its reference-rate procedure and publication occur at different times. A reference value is also an informational rate, not a transaction quote. A matching calendar date therefore establishes neither simultaneous observation with another market close nor immediate availability.
Keep the original record instead of silently filling gaps with a previous value. If a market is closed or a reading is delayed, mark it. Carrying a value forward may be appropriate for a specified task, but it creates no new observation and should not be counted as a fresh confirmation.
Write down what would weaken the explanation
For the teaching note, record three layers separately: the scheduled document and its new contents; the comparable yield observations proposed as a channel; and the GBP/USD change being explained. A dollar index reading is not automatically another independent vote, because it contains currency prices, including sterling. Its construction is a reason to inspect dependency, not a reason to ignore it in every analysis.
Now challenge the note. If corrected, aligned yields show the signed UK-minus-US spread increased rather than decreased, the simple “relative yields moved against sterling” description fails. If the currency decline began before the release, that release cannot explain the earlier segment. Neither finding predicts a rebound. They tell you which sentence to revise and which additional source to inspect.
Archive the cutoff, inputs, calculation, candidate mechanism and counterevidence together. Compare later cases with the same rules, including unresolved ones. This lets you learn whether a measure adds useful information without awarding points to several readings of one development. The full lesson develops dependence, synchronisation and correlation tests; the opening exercise first makes the small note auditable.
Three exercises with worked answers
The exercise data are hypothetical. Choose an answer, then check the calculation.
Build the full record behind a cross-market explanation
Continue to the dependency tree, shared-factor model, aligned series, documented 2024 episode and unresolved pre-meeting case. The detailed templates help distinguish a reproducible hypothesis from a story assembled after the outcome.
1. Why one signal is insufficient
Three model scenarios, without historical readings: VIX rises from 14 to 22, but EUR/USD remains near its previous price. Yields rise while the dollar weakens. At one broker, 80% of clients are long AUD/USD while the pair keeps rising. Each pattern has several possible explanations. The first may reflect US-equity-specific risk; the second term premium or relative debt repricing in both economies; the third client selection. Signs alone neither identify a cause nor automatically invalidate every earlier relationship.
One reading cannot establish an edge after costs without a separate test. Portfolio-strategy research documents momentum premia across asset classes and common factors behind their results[2], but studies portfolios spanning many markets over many years. It does not establish one measure's effectiveness on one currency pair — a different research question. A guaranteed “55–60% accuracy” for one signal without horizon, pair, costs and exit method is advertising, rather than research.
Informational advantage and trade performance are separate questions. A reading may improve the description of an event without reproducibly forecasting the next price change. Better forecasts, in turn, do not determine results after spread, slippage and financing. Here, we examine the first link: whether a new measurement adds to what earlier data already tells us. Trade construction and execution costs belong to the strategies section.
The instinctive answer is to add more measures. That creates this article's central problem. Another measure can add information about a different factor, or another measurement of the same factor with different error and delay. Mechanically duplicating the same series, however, creates no new evidence. Before adding a chart, specify its role and what the others do not measure: funding demand, a local shock, repriced protection or another participant group's response.
In Intermarket Analysis, Murphy (2004) developed an interpretative framework in which bonds, commodities and equities describe different cycle phases, and their interaction helps identify the regime[1]. This is a narrative framework, rather than measurable accuracy transferable to an individual market participant — any claim of “20–30 percentage points more accuracy” without methodology, horizon and exit method is unverifiable. Its useful practical implication is qualitative: layers should measure different things. MOVE and VIX can respond to the same shock, but describe volatility in different markets — US Treasuries and an equity index. Whether they add independent information in an episode must be checked, rather than assumed. Retail positioning measures a client subset; VIX measures the price of US-equity risk. Different definitions do not guarantee additional information for a particular pair.
The following four questions organise the analysis. They represent neither an empirically established optimal indicator count nor a condition for effectiveness. Each layer should answer a different question:
- Prerequisite: are the series and their availability times comparable?
- Layer 1: which hypothesis describes the regime and movement's source?
- Layer 2: do selected asset classes support the mechanism and add information?
- Layer 3: what do risk pricing and positioning show, and what remains unknowable?
Each layer serves one purpose: allowing a hypothesis unsupported by data to be rejected. The layer where it fails is informative — it identifies the missing mechanism link. “Unresolved” is a legitimate result, rather than a procedural failure — and a frequent one.
2. Four verification layers
Think of the procedure as successive filters, rather than a funnel leading to a decision. A filter retains what does not pass; its outcome is rejection, rather than selection. The four layers produce one of three hypothesis assessments: consistent, internally contradictory or unresolved.
| Layer | Question | Tools | Result |
|---|---|---|---|
| Prerequisite | Do all values describe a comparable window and were they available then? | Timestamp, time zone, calendar, price type, series version | Comparable / missing data |
| 1. Regime | Which hypothesis explains the change? | Relative rates, risk, demand and supply | Working hypothesis with a specified horizon |
| 2. Mechanism and measurements | What new information does each series add? | Yield differential, selected commodity, index or funding price | Agreement / contradiction / unresolved |
| 3. Risk pricing and positions | What is paid for protection and which client sample is described? | IV, risk reversal, an identified SSI reading | A limited conclusion about a field; not “whole-market consensus” |
Synchronisation precedes interpretation. Then assess the hypothesis, channel and readings that might contradict it. Agreement is not an end in itself. A fifth indicator measuring something distinct may change the assessment; one copying the first series adds no evidence. “Consistent” means agreement with the declared mechanism within the window, rather than a probability of price appreciation or a trading edge.
A key implication: “unresolved” is a legitimate procedural result, rather than failure. If the second layer diverges, do not seek a fifth reading to outvote it — record the hypothesis as unconfirmed and revisit the movement's source. This article's procedure ends there; managing exposure under an unconfirmed hypothesis belongs to the risk-management section.
Scroll the diagram horizontally or open it at full size.
2a. Four readings, one factor — recognising it
Start with a dashboard often treated as exemplary in market commentary. An analyst claims four independent observations confirm dollar appreciation: a rising dollar index, falling EUR/USD, a widening US–German bond-yield differential and higher pricing of the future US rate path. Four markets, four confirmations, high conviction.
In this simplified arrangement, all four observations may largely reflect one common impulse. Do not automatically count them as four independent confirmations — that requires investigation, rather than agreement alone.
Scroll the diagram horizontally or open it at full size.
Revising the dashboard — four observations become three questions
The revised set is structured differently, rather than enlarged. Instead of four readings from one branch, it uses three observations with different sources:
| Observation | What it concerns | Why it may add distinct information |
|---|---|---|
| Pricing of the future rate path | The underlying event itself | An observed contract price proxies expectations; it is itself a response to information and risk premia, rather than a direct measurement of the cause |
| The dollar against currencies outside the index basket | The move's reach beyond developed economies | Also responds to local factors and willingness to hold this asset class |
| The price of access to dollar funding | Demand for dollars as a settlement currency, rather than an asset | May contain a distinct funding-stress component; separate it from ordinary rate changes and quotation conventions |
Three observations with potentially different sources may contain more information than four closely related readings. More importantly, they can diverge — especially useful diagnostically. When every branch of one tree points the same way, the independent information need not match the chart count; when differently sourced observations disagree, the hypothesis needs resolution. Agreement can also be informative. Distinguish independent measurements, mechanical duplication and a common shock source — three separate questions.
Subtracting a common factor — a complete small model
Different ticker names are insufficient. The following MODEL contains eight synchronised one-day observations in hypothetical change units. They are neither historical returns nor a DXY estimate. Construct X=2F+A and Y=2F+B:
| t | F | A | B | X | Y |
|---|---|---|---|---|---|
| 1 | −1 | −1 | −1 | −3 | −3 |
| 2 | −1 | −1 | 1 | −3 | −1 |
| 3 | −1 | 1 | −1 | −1 | −3 |
| 4 | −1 | 1 | 1 | −1 | −1 |
| 5 | 1 | −1 | −1 | 1 | 1 |
| 6 | 1 | −1 | 1 | 1 | 3 |
| 7 | 1 | 1 | −1 | 3 | 1 |
| 8 | 1 | 1 | 1 | 3 | 3 |
Every mean is zero. ΣXY=32 and ΣX²=ΣY²=40, giving Pearson r=32/√(40×40)=0.80. Regressing either series on F gives slope ΣXF/ΣF²=16/8=2 and intercept zero. Subtracting 2F leaves residuals A and B. ΣAB=0 and ΣA²=ΣB²=8; residual correlation is 0. Agreement here comes from a common component we know because we constructed it.
In actual markets, choosing F is a hypothesis. A misspecified or omitted factor can leave an apparent relationship. Zero residual correlation proves neither statistical independence nor causality; positive correlation does not prove causality either. In a forecasting test, fit regression coefficients only on data available before the test period. Fitting the complete history and then “cleaning” earlier observations would use future information.
Correlation and beta answer different questions. Correlation is dimensionless and describes co-movement; regression beta describes the average change in Y units for one X unit in the sample. Doubling Y's scale leaves r unchanged but doubles beta. High correlation therefore does not specify the pair's response magnitude, and a shared chart axis does not replace calculation.
3. Layer 1: identifying the regime
A regime is a working description of conditions in which relationships are examined. Changes in rates, funding or premia may alter correlation's sign and strength, without following one fixed switching rule. Article 7.1 distinguishes stress sources; the categories below likewise require testing in a specified sample.
| Regime | DXY | Volatility | 2s10s curve | Commodities | Hypothesis requiring FX-market investigation |
|---|---|---|---|---|---|
| Risk-on | May weaken, or rise with US outperformance | May decline | Shape also depends on the policy path | Industrial demand may support prices | Distinguish broad risk appetite from one market's outperformance |
| Risk-off | Response depends on the stress source and funding | May rise; VIX concerns US equities | Falling short or long yields produce different slope changes | Gold may be sold for liquidity; oil also responds to supply | Check whether the funding currency, USD, CHF or only one of them gains |
| Reflation | Relative growth rates matter | No necessary sign | May steepen; separate 2Y and 10Y changes | Distinguish broad demand from supply disruptions | Do several exporter currencies respond similarly after accounting for rates? |
| Stagflation | No necessary sign | Risk pricing may increase | Policy and term-premium responses may oppose each other | Commodities need not all rise together | Import/export balances and central-bank responses distinguish currency reactions |
| Dash-for-cash | Check USD demand and funding prices | May rise sharply | Does not replace a liquidity measure | Selling may also affect gold | A broad currency move requires separate measurement of funding stress |
Before interpreting correlation, specify a mechanism hypothesis and sample-partition plan. Regime labels may be data-based, but selecting them after seeing results can rescue a hypothesis retrospectively. Imagine AUD and copper falling in the same window. One scenario is reducing positions funded in other currencies alongside a separate industrial-demand revision. Alternatively, both may respond to common growth news. Matching signs do not select the cause; distinguishing data is required. No coefficient or frequency is given for this example. Reconstructing correlation requires transformations, frequency, window, observation count and treatment of non-trading days, described in section 6.
Dollar Smile (article 7.2) is particularly useful here. The dollar may strengthen for very different reasons: US economic outperformance, or global stress and its safe haven (defensive demand). These two hypotheses require different measurements, despite the same direction of the dollar’s move.
Possible combinations to investigate. These model scenarios are neither necessary regime conditions nor an identification rule; they pose questions requiring the other economy and funding data:
- US-outperformance hypothesis: USD and US equities may rise with IV low relative to its own history. Examine relative growth and rate revisions, rather than US yields alone, which may also rise through inflation, supply or premia. AUD and NZD may gain against some currencies yet lose against USD; these pair signs do not prove the differential caused the move.
- Risk-reduction hypothesis: falling equities and higher VIX may coexist with gains in USD, JPY or CHF. Bonds and gold may rise through defensive demand, but may also be sold for cash. Neither rising gold nor identical signs for three currencies is necessary. Distinguish growth, inflation and funding shocks using rate, credit and currency-specific funding-cost data.
- Improvement-outside-the-US hypothesis: weaker USD and moderate VIX may be consistent with other economies becoming more attractive. Check both sides' data and rates. DXY can also fall through a euro-specific move without comparable changes in all currencies. EUR, GBP, AUD and NZD need not share a sign; local policy, trade, positions and funding can alter reactions.
4. Layer 2: confirmation across asset classes
Once you have a regime hypothesis, examine whether individual asset classes support it. Alongside agreement, ask whether it comes from one source — the procedure is in section 2a. Divergence can be particularly useful when contradicting a model forecast specified in advance. It can also reflect noise, a local shock or a data error. Its significance depends on the expected sign, horizon and tolerance recorded before assessment; agreement can also add information after controlling for a common factor.
Bonds
Debt-to-FX transmission mechanisms are described in article 7.3. Two matters are relevant here:
- Yield differentials. A narrowing US 10Y–Bund differential as US yields fall does not determine dollar direction. Two opposing channels may act simultaneously: a smaller differential means lower relative yields on those specific US 10-year securities versus German ones, while the same yield decline may reflect defensive demand supporting the dollar. The differential is neither currency-holding income, an OIS rate nor overnight funding cost; an FX swap needs a separate calculation. Until competing channels are separated, FX direction remains unresolved.
- Curve slope (2s10s). Inversion does not automatically confirm current risk reduction. It primarily describes the shape of rate-path pricing and the cycle phase — usually a longer horizon than current market sentiment. Combining it with same-day volatility mixes horizons into one assessment.
Commodities
The relationship between a commodity price and an exporter's currency is described in article 7.4. Here, use it only to check whether the commodity observation adds to the currency pair's own information:
- Oil (Brent/WTI): rising oil with broad risk appetite prompts checking whether CAD and NOK respond to a common demand factor. If supply restrictions drive oil, the relationship may differ — it does not confirm common demand, though it may add distinct information about the exporter's terms of trade.
- Gold: defensive demand and inflation concerns may support its price, while real yields, USD changes and liquidity needs can oppose them. Check the specified benchmark and window, rather than inferring a regime from the sign alone.
- Copper: is often treated as a global industrial-growth barometer. Rising copper with a steepening yield curve is sometimes read as reflation confirmation, but neither has one cause: copper may rise through supply factors, while debt supply or fiscal risk may steepen the curve. This is a hypothesis to investigate, rather than a rule.
Equity indices
Article 7.5 showed that different indices concern different currencies — with conditional, rather than mechanical, relationships.
- S&P 500 vs DXY. The relationship depends on whether growth, rates or liquidity stress dominate. In reflation, S&P 500 and DXY may diverge (risk-on, weak dollar); in a dash for cash they may move oppositely as equities are sold and dollars bought. Correlation has no single sign.
- Nikkei and USD/JPY are not mirror images.Prices do not directly reveal Japanese funds' FX exposures. Nikkei falls 3% and USD/JPY also falls — one hypothesis is unwinding yen-funded positions and repatriating capital. Nikkei falls 3% while USD/JPY rises — another is an external demand shock harming exports without repatriation. Signs alone cannot choose; actual flow evidence is needed. The distinction is fundamental and regime-dependent, rather than a simple correlation.
- DAX vs EUR. Some DAX companies earn revenue outside the euro area. FX affects revenue translation, but margins also depend on costs, hedges and sales markets. DAX may rise as EUR weakens; the sign alone does not identify exports as the channel. Specify whether a performance or price index is used. The claim is not “DAX correlates with EUR”, but “DAX and EUR respond to the same rate and growth regime, sometimes together, sometimes oppositely”.
Retail positioning
Retail-broker client positions show the side held by one narrowly defined group, rather than the market. What it measures, its denominator and why it is not the inverse of likely direction are explained in article 7.7. Here, ask whether it adds information absent from price itself.
5. Layer 3: is the hypothesis already priced in?
The third layer separates volatility pricing, relative protection pricing across the option wings, and a client sample's positions. They do not measure one “consensus”. High IV does not determine expected direction; RR shows neither all beliefs nor the complete options surface; SSI does not describe the whole market. Assess their value for a hypothesis by first establishing what each measures.
What the third layer can and cannot establish
IV describes a volatility quotation under a pricing model; alone, it does not determine a future exchange-rate sign. “High” requires context for the same option, tenor and convention. With other model inputs fixed, rising IV increases a typical vanilla premium, but actual-premium comparisons also require strike, spot or forward, time and discounting. A larger IV number does not prove every hedge costs more than usual, or provide a forecast of the actual movement range. Spot positions have separate cost components: spread, swap points and execution terms.
Risk-pricing information may change the uncertainty assessment without providing a new directional forecast. To claim a hypothesis “is already priced in”, identify an actual expectations variable and deviation from it, rather than merely high SSI or IV. Implications for exposure size and costs remain within the scope of section 9.
6. Series synchronisation — a prerequisite for a meaningful calculation
Before comparing two markets, ask: do both series cover the same time interval? Different windows can create apparent leading information. Not every timing difference necessarily does so.
Where apparent lead–lag comes from
The global FX trading week typically runs from Sunday evening to Friday evening in New York time. Equity, bond and commodity venues have their own opening and closing hours, holiday calendars and settlement times. Comparing a US close with an FX rate “from the same day” may actually compare readings several hours apart — systematically in the same direction.
If information arrives between cutoffs, the earlier close can create an apparent lag in one market. Not every timing difference does this; the result depends on information timing and price dynamics. After synchronisation, the relationship may disappear or persist. Compare both versions, preserve original timestamps and verify information availability before attributing the lead to market behaviour.
Four decisions to make explicitly
- Observation time. Use one common time for all series, with a time zone stated. If using the last available quote, specify an age tolerance in advance and flag stale readings. Quotes after the cutoff are inadmissible. Missing common measurements may require a different frequency or exclusion.
- Calendar. Align available observation times first, then calculate changes between common endpoints. Removing a day after calculating returns may leave a one-day change in one column and a two-day change in another. Carrying forward an old price creates no new quotation.
- Price type. Close, mid, central-bank reference rate and futures settlement are different series for the same instrument.
- Futures-series continuity. Splicing successive contracts can create roll-date jumps that are not market movements. Without the roll methodology, the calculation partly measures how the series was assembled.
A holiday: the same row need not represent the same window
MODEL: X prices are 100 on Monday, 102 on Tuesday and 101 on Wednesday. Y is 50 on Monday, its market closes Tuesday, and it is 51 Wednesday. Carrying 50 forward creates an artificial 0% Y return alongside +2% X. Wednesday's Y return of +2% covers two days, while X=(101/102−1)×100=−0.98% covers one.
For common Monday–Wednesday movement, first select 100→101 and 50→51. This gives +1% and +2% over the same two-day window. One common observation cannot yield a meaningful correlation. In a larger sample, each row needs matching time boundaries; differing durations must be flagged or separated according to the plan.
One r can conceal two opposing periods
The next MODEL contains 12 consecutive, synchronised one-day percentage returns, without holidays or gaps. X=(−3,−2,−1,1,2,3,−3,−2,−1,1,2,3); Y=(−3,−2,−1,1,2,3,3,2,1,−1,−2,−3). Calculate Pearson correlation on returns, without accumulating levels or lagging the series.
| Range | N | r | Conclusion for the model only |
|---|---|---|---|
| Full sample 1–12 | 12 | 0 | Averaging conceals the sign change |
| 1–6 | 6 | +1 | First subperiod: matching changes |
| 7–12 | 6 | −1 | Second subperiod: opposite changes |
| Window 4, endpoints 4–12 | 4 in each | 1;1;1;−0.232;0;0.232;−1;−1;−1 | A short window spanning the boundary can also fluctuate |
The formula is Σ[(X−meanX)(Y−meanY)] / √[Σ(X−meanX)²×Σ(Y−meanY)²]. Window 4 was chosen only for this short exercise, rather than as a market recommendation. Successive windows share observations, so nine rolling-chart points are not nine independent tests. In market research, specify window length and subperiods before results; relabelling regimes after examining r may conceal hypothesis-fitting.
One observation and a small N can change the entire conclusion
Append a thirteenth pair to the same 12: X=30%, Y=30%. Pearson rises from zero to 0.937. This does not discover a stable relationship; one shared extreme point changed the calculation. In real data, first check quotation errors, units and the market event. Do not remove a valid crisis reading just because it disrupts the result. Show calculations with and without it, and a method less sensitive to magnitude, such as Spearman rank correlation. None replaces justification of sample selection.
UNCERTAINTY MODEL: with r=0.50 and independent bivariate-normal observations, Fisher's transform z=atanh(r), standard error 1/√(N−3), gives an approximate 95% interval of −0.189 to 0.859 for N=10. For N=60 it is 0.282–0.669. Calculate the bounds as tanh[z±1.96/√(N−3)].[15] This demonstrates assumptions, rather than an interval for a particular pair. Autocorrelation, volatility and overlapping horizons undermine treating financial rows as independent; a method preserving temporal dependence, such as block resampling with an explicit block length, is then required.
7. An intermarket observation dashboard — what to display
The dashboard should show sources and data roles. Free or paid access does not determine measurement independence: two screens may use the same series. Options data and order-book detail vary in access; possessing a terminal alone does not reveal complete dealer positioning. Record each selected series' provider, convention and available timestamp.
The question defines the minimum; the catalogue is optional
For relative US/euro-area policy changes, a minimum dashboard may contain three raw series: EUR/USD and comparable US and German 2Y yields. Form one differential from the two yields; do not count it as a fourth independent confirmation. If funding stress is a competing explanation, add an identified stress measure. DXY checks basket breadth, but its euro component prevents it from being a separate EUR/USD counterpoint. There is no universal minimum number of independent markets: distinguish raw calculation inputs from separate transmission channels.
The following catalogue assists selection, rather than mandating eight confirmations. Verify platform symbols with the provider; a ticker alone does not define methodology.
| Ticker (TradingView) | Instrument | What it tells you |
|---|---|---|
| DXY | ICE dollar index (six currencies) | Dollar conditions against developed-market currencies. Not a broad global dollar index — it includes neither yuan nor any emerging-market currency[9] |
| US 10Y | US bond yield, 10 years | Expected short-rate path plus term premium — the chart alone cannot separate them, and they respond to different factors |
| TVC:VIX | Cboe Volatility Index | Expected 30-day S&P 500 volatility implied by SPX options; an indirect measure of US-equity risk pricing, rather than a “mood” measure |
| ES1! | S&P 500 futures (continuous series) | Trades almost around the clock; a continuous series includes roll jumps and differs from cash SPX |
| XAUUSD | Spot gold | Safe haven vs inflation hedge |
| WTI oil contract | An explicitly selected contract, rather than a generic ticker | Global demand and exporter currencies. Labels such as “USOIL” are provider-dependent and define neither the contract nor rolling methodology — different services may show different series under the same name |
| EURUSD | EUR/USD | Major pair; its euro component contributes mechanically to DXY, without reproducing the entire basket |
| AUDJPY | AUD/JPY | Proxy for willingness to hold interest-rate-differential positions |
Organise by roles, rather than window count
First display the pair and quotation convention: rising EUR/USD means EUR appreciates against USD. Beside it, show the mechanism measure, such as changes in comparable 2Y differentials in basis points. Add a reading that could contradict the hypothesis and a metadata card: price type, tenor, time zone, calendar and information available at the cutoff. S&P 500 price, total-return and continuous-futures series are not interchangeable. Feed-specific spot gold, LBMA fixing and futures likewise need separate specifications; WTI and Brent are different benchmarks. Platform-interface settings are described in the tools-and-platforms section.
An additional input: retail sentiment
- IG Client Sentiment (free on IG's website) — shows IG clients' long/short percentages for major pairs. The provider describes continuous updating as positions open and close; visible publication frequency and delay require product-specific checks.
- OANDA Order Book (free on OANDA's website) — shows open orders and positions at this single broker. Orders concentrated at a level are not necessarily protective stops: opening and profit-taking orders can occupy it too, and the book does not distinguish their purposes.
- Myfxbook Outlook — aggregates positions across several brokers.
8. Case study: July–August 2024, unwinding interest-rate-differential positions
Analytical question. Did equity, volatility and rate observations in early August 2024 contain information about AUD/JPY absent from the pair itself? Horizon: several sessions. The episode is examined retrospectively solely to show where the procedure would apply and where it would not — rather than reconstructing a live decision.
Chronology and measurement sources
On 31 July 2024, the BoJ announced an overnight target of approximately 0.25%, effective 1 August.[8] The BIS analysis of 27 August describes the earlier yen turn, responses to US data and subsequent intensified deleveraging.[11] The entire episode is therefore attributed neither to one decision nor one channel. This is the authors' retrospective interpretation, rather than live observation of all participants' positions.
| Date in 2024 | ECB EUR/JPY | ECB EUR/AUD | AUD/JPY cross | Cboe VIX close |
|---|---|---|---|---|
| 3 July | 174.18 | 1.6123 | 108.03 | Not required for this window |
| 31 July | 162.76 | 1.6635 | 97.84 | 16.36 |
| 2 August | 161.37 | 1.6629 | 97.04 | 23.39 |
| 5 August | 155.98 | 1.7008 | 91.71 | 38.57 |
| 9 August | 160.33 | 1.6609 | 96.53 | 20.37 |
The cross uses daily ECB reference values: divide JPY per EUR by AUD per EUR.[12] Without rounding intermediate values, the 31 July→5 August change is −6.27%. The complete ECB historical XML was retrieved, without filling gaps. This is a European-procedure reference, rather than a Tokyo close or traded low. VIX comes from the administrator's daily CSV; its US close occurs at a different time. The table documents the episode's development, rather than forming an AUD/JPY–VIX correlation sample.
In the same Cboe CSV, the 5 August VIX high is 65.73, and the close 38.57.[7] The difference is 27.16 index points; substituting the high for the close would overstate the level by 70.42% relative to the close. The daily row does not state the time of the high. Do not therefore treat 65.73 as information known before a particular earlier moment, or equate VIX with VIX-futures prices or AUD/JPY volatility.
What each observation adds
Source and time: the BoJ decision, earlier yen movement and subsequent US data are distinct chronological elements. Belonging to one episode does not make them one shock. FX and equity prices may share a risk-reduction component while also reflecting local responses. Without a model and synchronised data, no independent-confirmation count or percentage causal contribution is given.
Protection pricing: this reconstruction includes no AUD/JPY IV and risk-reversal archive with aligned tenor and timestamps. It cannot claim that options warned, or that no warning existed. An unestablished field remains unestablished; it is not converted into no market response.
Synchronisation: ECB daily values, Tokyo closes and US closes cover different parts of the day. Investigating reaction order requires common intraday cutoffs and data-release times. This material checks daily chronology and units; it establishes neither lead–lag nor the procedure's advantage.
9. An unresolved example: readings before a central bank meeting
Qualification.The values in this example are illustrative — they come from no actual meeting and are not historical records. They show a configuration ending with “unresolved” and why that is a legitimate outcome, rather than failure.
The readings
Model hypothesis: volatility pricing rose in the selected options cross-section without pronounced 25-delta asymmetry. This does not mean all participants “have no view” or that the physical distribution of future exchange rates is known. Exercise horizon: T₀ to T₁ before a scheduled meeting. Both readings are fictional and known exactly at the cutoff; model delay is zero.
Skew here means the one-month 25-delta risk reversal: the implied-volatility difference between EUR calls and puts at the same absolute delta, with positive sign indicating higher EUR-call IV. It compares volatility quotations, rather than monetary option premiums directly. The model uses unadjusted spot delta; both options concern EUR, 1M tenor remains fixed, and values are annualised volatility points. Actual quotations must retain precise expiries and delta conventions: different maturities may contain different events.
The assessment rule precedes the numbers
This score organises the description without calibrating probabilities or effectiveness. Predeclare three equally weighted fields: ATM IV rises at least one volatility point; |25Δ RR|≤0.10 points; the absolute US–German 2Y differential change is below 5 bp. A condition met scores 1, unmet 0, missing data “?”. Do not count DXY and EUR/USD as extra separate votes. Assessment applies to this narrow model; 3/3 is neither a signal nor three independent proofs.
| Model series; common T₀/T₁ | T₀ | T₁ | Transformation |
|---|---|---|---|
| EUR/USD mid, USD per EUR | 1.1000 | 1.1000 | Return 0% |
| US sovereign 2Y yield %, same convention | 4.00 | 4.01 | +1 bp |
| German sovereign 2Y yield %, comparable tenor | 2.00 | 2.01 | +1 bp; differential unchanged at 200 bp |
| EUR/USD 1M ATM IV, annualised % | 8.00 | 10.00 | +2 volatility points |
| EUR call +25Δ, unadjusted spot delta, 1M IV % | 8.00 | 10.00 | Same tenor and convention |
| EUR put −25Δ, unadjusted spot delta, 1M IV % | 8.00 | 10.00 | RR=call−put=0 at both times |
The model scores 3/3 known fields meeting the definition. Higher volatility pricing and no asymmetry at these two points describe option prices; rates show no relative repricing. The future EUR/USD direction remains unresolved. Even an identical score in another window would not establish independence or stability beyond this exercise. Thresholds have no empirical validation and are not entry rules.
The four layers
Source. The source event is a scheduled meeting — its date is known ahead of time. This distinguishes it from the episode in section 8, where the source was identified afterwards. The implication is less intuitive: the known date lets us choose an event-containing tenor, but does not establish that its contents were predicted correctly; price may respond to deviation from pre-release expectations. The date, quotation convention, turnover and observation delay may also provide useful information, distinct from a surprise in the decision's content.
Channel. Rising IV and spot consolidation may share anticipation of a scheduled event, but are different measurements: the former describes future uncertainty pricing derived from options, the latter the spot path already realised. Before counting them as independent evidence, check whether they reflect the same expectation — see section 2a.
Is it already priced in? RR near zero means similar call and put IV at the two selected 25-delta points for the given tenor and convention. It does not mean equal monetary premiums: strikes, discount factors and pricing terms may differ. Nor does it establish symmetry of the full surface or absence of all directional information. The narrow |RR|≤0.10 condition is met here; these two quotes and this score do not forecast EUR/USD's future sign.
Synchronisation. IV is quoted OTC, bond yields have their own cutoffs, and FX trades continuously. Comparing all three “from the same day” requires one observation time — without it, even their description of the same state is uncertain.
Why this is a result rather than no result
The procedure ends here without a directional conclusion — which is the correct outcome. The readings are not contradictory: they do not identify the direction of the future change. The data support “volatility pricing increased”, but not “the direction of the meeting’s outcome is known”. Distinguishing these two statements is the entire value of the example.
10. Common mistakes
a) Correlation ≠ causality
AUD and iron ore may respond to common industrial demand, while a higher export-commodity price may also alter terms of trade and expected exporter revenue. Thus absence of every direct channel cannot be asserted. Correlation does not identify the dominant mechanism. Supply-driven and demand-driven iron-ore increases may produce different FX responses; distinguish them using shock source, benchmark and horizon. Shifting a chart by one day does not identify causality.
b) Overfitting the rule
Combining conditions may reduce the episode count, but N cannot be guessed in advance. Count it and distinguish successive days of one state from independent events. TEST-COUNT MODEL: 20 pairs×3 windows×5 lags give 300 variants before thresholds are selected. The best result requires different assessment from one predeclared test. Record unsuccessful variants too, freeze the model and reserve later data for evaluation without further tuning. Extra parameters may improve fit, but forecasting effects depend on data and regularisation; there is no law that every added parameter always worsens future results.
c) Cherry-picking
Episodes with clear outcomes are more memorable than contradictory readings yielding no conclusion. This is selective memory, described in cognitive-heuristics literature[6], rather than a market phenomenon. The hypothesis-diary template below addresses it: record every hypothesis, including “unresolved” outcomes — precisely the cases memory tends to omit.
d) An expanding measurement set
Four layers gradually expand: a fifth measure, then a sixth, then a dozen after a quarter, with one always contradicting the others. The criterion is added measurement value beyond existing knowledge. A new reading may add another aspect, improve precision or merely duplicate earlier data. Duplication may increase conviction disproportionately to information; the check is described in section 2a.
e) Ignoring regime changes
Link strength need not remain constant. Market-integration research describes changing interdependence, rather than replacing a test for the actual currency pair.[3] Forbes and Rigobon discuss conditional correlation in selected equity-crisis episodes: increased volatility can raise an unadjusted coefficient without stronger transmission.[4] This is not universal proof that every correlation increase is erroneous. Partition the sample according to the plan, inspect extremes and methodology changes, and treat a new regime label as a hypothesis to evaluate. Low full-sample correlation likewise does not prove no relationship: the model in section 6 contains opposing subperiods.
How to guard against these errors
A practical procedure:
- Keep a hypothesis diary — distinct from a trading journal. Record every intermarket hypothesis, including those followed by nothing. The template is below. Accumulated records allow reliable calculation of matters poorly reconstructed from memory: the share of unresolved hypotheses and readings most often found dependent on each other.
- Review the calculation quarterly. Overlaying charts is not a correlation audit — it shows only shape similarity at software-selected scales. A minimum assessment requires the same transformation (preferably returns, rather than levels), common observation times and window, rolling correlation instead of one number, observation counts and subperiod comparisons. See section 6.
- Use a “one-screen” rule. If the dashboard does not fit one monitor, remove the weakest instrument. Measurement sets expand gradually and unnoticed — each addition looks justified, while their sum produces a screen from which no conclusion can be drawn.
- Treat later-data testing as one check, rather than a final verdict. Results similar to historical testing do not prove robustness: 30 days may cover one regime and very few independent cases. A test's significance depends on independent-observation count, regime coverage, included costs and whether rules were frozen before the test began. Without these four conditions, it is another historical test on newer data.
Hypothesis-diary template
Seven fields. Make the record when the hypothesis is formulated, rather than after the outcome is known: recording expectations beforehand limits fitting them to a familiar historical outcome. In this procedure, these fields are not filled retrospectively; historical reconstruction is labelled separately.
| Field | What to enter | Example |
|---|---|---|
| Timestamp | Date and time with a time zone. Without the zone, later series comparisons are unusable — see section 6. | MODEL: T₀=14:30 UTC; every value known before the cutoff |
| Hypothesis | One declarative sentence about the movement's source, rather than price direction. | Dollar appreciation over the last three sessions reflects repricing of the US rate path. |
| Mechanism | Transmission from the source to the observed market. If it cannot be specified, the hypothesis is a correlation without substance. | Higher rate-path pricing → higher short-end yields → wider rate differential versus the euro area. |
| Expected confirmations | Observations expected if the hypothesis holds, distinct from the hypothesis's own measurement. | Higher US two-year yields; weaker responses in pairs unexposed to the rate differential. |
| Contradictory signals | Observations already contradicting the hypothesis. A blank field warns about the author, rather than the market. | Gold rises alongside the dollar — more consistent with defensive demand. |
| Horizon | The window in which the hypothesis applies. Fix it before results, rather than extending the wait after a forecast fails. | The next two common 14:30→14:30 UTC windows, under the specified calendar. |
| Invalidation condition | An event after which the hypothesis is rejected, specified so it can be checked without interpretation. | MODEL: the comparable 2Y differential returns to its pre-T₀ value while EUR/USD movement persists in both declared windows. Assess any competing channel separately. |
Also record the data version and benchmark: which values were available, when the file was retrieved, whether it was subsequently revised, what simpler description uses price alone, and how many variants were previously tested. If horizons overlap, the boundary between the training and test sets must exclude shared future windows; otherwise, the same movement leaks into both parts of the sample. Assess the additional measure against the same benchmark, rather than by comparing two different periods.
After the horizon closes, add an eighth field: assessment — consistent, internally contradictory or unresolved. Use unresolved when data is missing or competing mechanisms fit the observation. If almost all entries are “consistent” or “contradictory”, check whether criteria force a verdict despite uncertainty. The assessment proportions alone do not prove author error.
For the next review, choose one pair and one mechanism question. Before seeing results, record window boundaries, each raw series' role and an observation that would change the assessment. Missing timestamps or required measurements leave the result unresolved. The next section turns to strategies, beginning with the distinction between style and system.
FAQ — intermarket analysis in practice
Do I need Bloomberg for intermarket analysis?
Public sources often suffice for definitions, daily series and simple calculations. A paid terminal can improve access and timing precision without guaranteeing complete dealer positioning or independent data. The provider, convention, coverage and actual reading's timestamp matter more than the tool's name. If a required series is unavailable, record the limitation rather than reconstructing it from another indicator and counting it as a separate measurement.
When does correlation break down — and what should I do?
Correlation can change with conditions, but one divergence does not establish a new regime. Copper rising while AUD falls may reflect ordinary volatility, a local shock, policy changes, supply disruption or data/window errors. Comparable observations are needed to evaluate the predeclared expectation. This procedure calls for rechecking the movement's source, comparing horizons and verifying synchronisation, then recording an unresolved outcome. Divergence is no mean-reversion trading signal either: it describes knowledge, rather than price direction.
Does a regime cause correlation?
Not strictly. “Regime” summarises factors — central-bank stance, liquidity pricing, commodity-price direction and willingness to hold leveraged positions — rather than one observable, measurable cause. The label usefully shortens description, but misleads when replacing the mechanism question: “because of risk-off” sounds explanatory while merely naming what remains to be explained. This procedure requires a specific transmission channel, rather than a label.
How long does a daily intermarket review take?
Time depends on data readiness, hypotheses checked and divergences encountered. One might reserve a few minutes for a fixed checklist and a separate block for investigating causes — an organisational example, rather than a measured speed standard. Record a cutoff and time zone for every review. If the short block cannot verify data quality or the mechanism, mark the outcome unresolved instead of forcing an explanation.
Which intermarket-analysis errors are most common among beginners?
Pitfalls discussed here: (1) confusing correlation with causality — AUD and iron-ore co-movement does not identify the dominant channel; (2) overfitting — building ten-parameter rules from 15 observations; (3) expanding the measurement set until analysis stalls; (4) ignoring regime changes — using a relationship from six months ago that no longer holds; (5) keeping no diary — remembering successes and forgetting failures, distorting accuracy. Another is confusing a persuasive narrative with an edge. Assessment here relies on sources, a proposed transmission channel, each measurement's role and a condition weakening the hypothesis. Without those checks, a coherent account may be interesting but does not prove a forecasting advantage.
Is there a ready-made indicator combining cross-asset signals?
Yes. The Chicago Fed NFCI is a weekly financial-conditions index covering money, debt and equity markets and the banking system.[13] Its aggregation method does not produce a universal FX signal. Do not confuse it with monthly economic-activity index CFNAI. A custom score is possible too, but weights, thresholds, data state and evaluation must precede testing. This lesson's 3/3 model organises description without calibrating probabilities.
Sources and bibliography
Academic research
- Murphy, J.J., Intermarket Analysis: Profiting from Global Market Relationships, Wiley, 2004 — a classic account of equity, bond, commodity and currency relationships; a four-asset-class interpretative framework. Background bibliography, rather than a test of this procedure's effectiveness.
- Asness, C.S., Moskowitz, T.J., Pedersen, L.H., “Value and Momentum Everywhere”, Journal of Finance, Vol. 68, No. 3, 2013 — author description and AQR abstract; coexisting value and momentum premia across eight asset classes and their common factors. Portfolio behaviour, rather than the number of dashboard measures.
- Bekaert, G., Harvey, C.R., “Time-Varying World Market Integration”, Journal of Finance, Vol. 50, No. 2, 1995; subsequent work: Bekaert, Harvey, Ng, “Market Integration and Contagion”, Journal of Business, 2005 — author's 1995 abstract; capital-market integration, rather than a direct FX-pair test.
- Forbes, K.J., Rigobon, R., “No Contagion, Only Interdependence: Measuring Stock Market Comovements”, Journal of Finance, Vol. 57, No. 5, 2002 — NBER WP 7267, 1999 version and 2002 publication description; abstract reviewed. Conditional correlation in the equity crises studied, rather than a claim about every FX correlation increase.
- Ilmanen, A., Expected Returns: An Investor’s Guide to Harvesting Market Rewards, Wiley, 2011 — publisher metadata and description; background on assessing expected returns. The complete book was not read, nor used to derive the lesson's model test count.
- Tversky, A., Kahneman, D., “Judgment under Uncertainty: Heuristics and Biases”, Science, Vol. 185, No. 4157, 1974 — availability heuristics and selectively recalling confirming cases.
Institutional documentation and data series
- Cboe Global Markets, Cboe Volatility Index (VIX) — White Paper: Cboe Volatility Index Methodology and historical series VIX Historical Data (daily open, high, low and close), cboe.com/tradable_products/vix/ — historical data and complete daily CSV, retrieved 4 October 2026. VIX describes expected 30-day S&P 500 volatility implied by SPX options, rather than “mood” or the cost of any arbitrary hedge. The 5 August 2024 intraday high comes from the index administrator's series; close 38.57 comes from its CSV, with high 65.73 recorded in a separate HIGH field.
- Bank of Japan, Statement on Monetary Policy, 31 July 2024 — raising the uncollateralised overnight target to approximately 0.25%; and Changes in the Monetary Policy Framework, 19 March 2024 — ending negative rates and yield-curve control. July document and March document; decisions and effective dates reviewed.
- ICE Data Indices, U.S. Dollar Index (USDX) — Contract Specifications and Index Methodology, official ICE FAQ, definitions and weights reviewed, accessed 4 October 2026. Geometric construction, normalising constant 50.14348112 and six currency weights, including euro 57.6%. Further discussion: article 7.2.
- Swiss National Bank, Swiss National Bank discontinues minimum exchange rate and lowers interest rate to −0.75%, press release of 15 January 2015, full SNB announcement, reviewed 4 October 2026.
Reconstruction and calculation sources
- Aquilina, M. et al., BIS Bulletin 90, 27 August 2024 — chronology and amplified deleveraging description reviewed; the authors' retrospective interpretation, rather than proof of one source for the entire episode.
- ECB, reference-rate methodology and publication and complete XML, retrieved 4 October 2026. AUD/JPY=EUR/JPY÷EUR/AUD for five specified dates; neither a traded price nor a series synchronised with VIX closes.
- Federal Reserve Bank of Chicago, About the NFCI — index coverage, methodology and frequency reviewed; distinct from CFNAI.
- BIS, Global FX markets: when hedging takes centre stage, December 2025 — April-episode analysis and interpretation of USD hedging demand reviewed, without generalising to every stress period.
- R Core Team, cor.test: Tests for association/correlation — documentation of the approximate Fisher interval for Pearson correlation and its assumptions. The lesson's eight- and twelve-observation models have complete inputs and calculations, rather than historical market data.