GlobalFlows

The idea

Read left to right. That is the order money actually travels: cash conditions show up first — the dollar votes here — then rates pass the price of money around, growth and inflation say what households and firms are doing, and risk tells you whether markets have noticed yet. When you know that mix, you can ask the useful question: so what for assets?

1. Liquiditycash · dollar · plumbing
2. Ratesyields
3. Growthreal activity
4. Inflationprice pressure
5. Riskfear & spreads

Tap a component — or All, FX, Markets — to jump to that section of the book. Compare, Chart, and the lookback stay put. Markets is grouped with titled rows (Duration, Credit, Equities, and the rest), not a second row of tabs.

Where the numbers come from

Public sources: FRED, the New York Fed, Yahoo Finance, Bundesbank, Bank of England, Japan’s Ministry of Finance, S&P earnings yield from multpl, and the CFTC gold positioning report (via FuturesBench). China’s credit stock is the BIS series, carried on FRED. Every row names its own source when you tap it. A morning job at 8:47 New York time — after the 8:30 prints, before the open — pulls the book, rebuilds the archive from 2003, refits the calibration, writes today’s regime, then runs the sanity check. If the check fails, that file does not publish.

Market prices move every day; payrolls arrive monthly; GDP and China’s credit stock arrive quarterly and late. A series that has gone quiet past its allowed age stops voting rather than casting a stale opinion. It stays on the book until it is months old, then it is hidden. Methods and today’s fitted centres are on Math.

Regime components

Each component answers one question: is cash cheap or scarce, is the price of money easy or tight, is real activity strong or soft, is price pressure hot or cold, is the market calm or frightened.

• Chevron — ▲1m up, ▼1m down, –1m flat. Which way it has moved over one month. This is the turn, not the level. The table lookback does not move it.
• Reading — Five states. Full green or red means it has crossed ±0.45, and the reading is the full name (Hot, Tight, Strong). A lean paints the needle dim green or dim red. White is only Mid / Neutral. Never full green with Mid, never white with Hot. The lean is a mirror: halfway from 0 to ±0.45. Liquidity Easing / Leaning easy / Neutral / Leaning tight / Tightening; Rates Easy / Leaning easy / Neutral / Leaning tight / Tight; Growth Strong / Leaning strong / Mid / Leaning soft / Soft; Inflation Hot / Leaning hot / Mid / Leaning cold / Cold; Risk Risk-on / Leaning risk-on / Neutral / Leaning risk-off / Risk-off.
• Score and needle — −1.5 to +1.5. Full green and red take the ends at ±0.45. A lean uses the same hues, dimmer. White is Mid / Neutral. The needle is drawn on ±1; past that the number can still climb. Positive is the reflationary end. If the indicators behind it disagree sharply, the tap says so. The reading still names the average.

All five share one axis. Green is the reflationary end — more cash, easier funding, firmer growth, hotter prices, more appetite for risk. Red is the contractionary end — the same five things the other way. White is Mid / Neutral. A lean is dim green or dim red — same family, not the call. Mixed belongs to the asset classes. Neither end means good or bad. Green on Inflation is Hot: prices are running, not the app approving. That is the one trap. On the six asset classes, green is in favor and red is out of favor. Mixed is the reading and the needle lean, not a third colour.

1w · 2w · 1m · 3m · 6m is the lookback for the table and charts. Three months is the table default. Chevrons, duration, credit, and in / mixed / out stay on the 1m turn. The table buttons do not rewrite the six classes.

Dollars are made in more than one place. Liquidity watches four balance sheets, not only the Fed: the Fed, the ECB, the Bank of Japan, and China. The app calls that set G4 — note it is not the usual G4, because the Bank of England is not in it and the China leg is FX reserves rather than the PBoC total, which FRED does not carry. Reserves are the part that recycles into world markets. Rates includes German, UK, and Japanese 10-years. MOVE can turn Rates Tight, never Easy — calm bond vol does not vote Easy. Growth and Inflation are still US reads. China’s credit stock is shown but does not vote: it is a stock that only ratchets upward, it has been at the top of its range for years, and BIS publishes it two to three quarters late. A number that cannot move is not a signal. China reaches Liquidity through FX reserves in that G4 group instead.

Growth’s coincident group is jobs, claims, GDP, and the activity indexes, and it carries half weight. That is on purpose. Jobs and GDP describe the quarter that just ended; the Empire and Philadelphia factory surveys describe the one starting. When the surveys are at the rail while the hard data is still Mid, the hard data has moved their way within a quarter about two thirds of the time since 2003, against a coin flip on every other day. So the surveys may slide the needle — and when they are out ahead like that, the tap says so. The reading is early, not confirmed.

The six asset classes

Under the five sit six asset classes: Treasuries (5 / 10 / 30), Credit (IG / HY), Equities (Cy / Df), Crypto, Gold, Commodity (Oi / Cu). Green is in favor, red is out. Mixed is the reading, the needle lean, and the tap — not a third colour. The needle shows lean within the call. It cannot flip the call. The table lookback does not rewrite in, mixed, or out.

The six read the score on each component, not only the reading on the box. A score counts as fully reflationary at the colour line +0.45 and fully contractionary at −0.45, so Risk at +0.40 is leaning risk-on for every net: fear still looks cheap. Easy and tight weights are mirrors on that same rail: 0 at the center, full at ±0.45. Neutral is quiet on both sides. A mix that looks late — Growth Strong or leaning strong while fear is still cheap — takes Equities and copper out. High yield is the other side of that mix: it outs when growth is soft into cheap fear. Over the full 2003–2026 record those overlays have an edge on about a one-month horizon. The years since 2023 are a thinner sample: Credit almost never got to in, Gold almost never got to out.

• The pair line follows the six when they have already chosen. Equities out and long Treasuries in is long bonds over stocks — even if duration risk is mixed (that mix is 5s versus 10s, not silence).
• Duration risk rising and credit risk falling → stocks over long Treasuries
• Duration risk falling and credit risk rising → long Treasuries over stocks
• Both rising, cash paid → cash over stocks and bonds
• Both falling → risk assets and duration can both work

• Treasuries — The 5s are cash (the Rates call). 10 and 30 are duration. The parent call and the parent needle are the long end only; the 5s are reported next to them, not voted into them.
• Credit — IG is duration plus credit. HY is growth and fear, and whether spreads pay you.
• Equities — fear, growth, and whether you are paid. Strong or leaning-strong growth while fear is still cheap is late. Cy leans growth; Df leans fear.
• Crypto — Bitcoin. Follows Liquidity: easy plumbing is the bid; a drain or a lean that way is out, not a bounce. Does not vote a component.
• Gold — crisis plumbing, an inflation wage, or deflation fear. Crowded speculative longs tax the box. Sometimes none of those is live, and the app says so; sometimes two are live and they cancel.
• Commodity — oil carries the dollar and inflation; copper wants paid fear or soft growth, not firm growth into cheap fear. They are not the same trade.

Tap a class for why and the split. History sits under that: one verdict for the class, then a short rose-in line for each tenor or sleeve (5s / 10s / 30s, IG / HY, Cy / Df). Gold and Crypto stay one full comparable — they have no split. Regime is the longer sheet: the same class history, then Watch — the print that would prove the call wrong.

What happened last time

The app finds the days in its 23-year archive whose five-component mix most resembles today, and reports what happened next: the median move and how often it rose. A base rate, not a forecast.

On a split class the verdict is stated once. Each sleeve keeps its own rose-in rate, because 5s, 10s, and 30s do not move as one. A close match is worth weight; a distant match is context. When history disagrees with today’s call, the tap says so. The class title does not change colour.

Two histories, and only one of them is a record

This is the distinction to hold on to, because everything on the site that cites the past comes from one of two places and they are not worth the same.

The archive is 2003 to today, about 6,000 days. It is a replay, not a track record. The app did not exist in 2003. The archive is today’s rules run backwards over data as it reads now, including figures revised long after the fact, and it is rebuilt from scratch every morning. Change a rule and the whole 23 years changes with it — the day coincident was halved so Growth surveys could slide the needle, all 6,000 days were recomputed. Every base rate and every one of the standing checks below stands on it. It is the best available way to ask whether the rules are sane, and it can never be evidence that they work, because it already knows how the story ended.

The forward record is the other one. Each morning the six calls are written down and never touched again, then marked when the returns arrive. It began in September 2026 and currently holds a handful of days. That is the only test the app cannot flatter, and it is nearly empty — which is the honest state of it, not a placeholder.

Where the two are mixed up, the backtest always looks better. So: the archive says a mix like today’s has historically paid this. The forward record says what this app actually called, in advance, and how that went. Only the second one is a claim about the app.

How we know it is right

Nothing here has ever broken loudly. Every real bug ran clean, produced a plausible number, and stated something false with total confidence. Five standing checks, plus the forward record. Three of them read the archive, so they can catch rules that are broken but cannot prove rules that work. The morning job will not publish if the first check fails.

• Sanity — recomputes all eleven boxes independently and confirms they match what was published.
• Plumbing audit — the machinery rather than the rules. That today’s reading and the 23-year replay are scored by one model and not two; that a band edited in the code actually reached the data; that every series prints with a scale; that the forward record holds no day that has not happened; and that the numbers quoted on this page and on Math are still the numbers in the code. Each of those has gone wrong at least once, silently.
• Band audit — flags a series pinned at an extreme, or a colour that almost never fires.
• Episode audit — lines the five up against 24 named events, each with an expected reading written from what actually happened.
• Call audit — whether classes called in favor actually outperformed the ones called out, counted in independent windows, not one episode counted fifty times. Still the archive, so still a backtest.

The distinction the five rest on: Rates reads Easy through the 2013 taper tantrum. The 10-year went from 1.6% to 3.0% in four months — that looks wrong until you notice the 2-year sat at 0.2%, QE was still running, and the curve steepened. Rates were easy the whole way; they became less easy. The drama lives in the turn. A reading that went red there would be reporting the turn and calling it the level.

The only test that cannot flatter the app is forward time, and that is the forward record described above. It prints empty on purpose until a class has twelve independent windows, and it is not on this site yet — it runs alongside the other three checks and will appear here once there is something in it worth reading. Today it holds weeks of overlapping days, which is nothing.

Honest limits

• The archive starts in 2003. It has never seen the 1970s, and it has never seen Volcker.
• Some economic vintages only reach back to around 2016; before that the app is reading revised figures.
• Bitcoin’s history starts in 2014. Its base rates are not an equal sample beside Treasuries or equities.
• “In favor” means this mix has historically rewarded owning this, over about a one-month horizon, in the 2003–2026 archive. Since 2023, Credit has almost never printed in and Gold has almost never printed out — this cycle is not the whole archive. A base rate, not a target, and not advice.
• Bands and checklists were fitted on the 2003–2026 archive and graded on the same 2003–2026 archive. The colour cuts are set from that archive’s own spread, so it cannot independently confirm them. The exact weights are judgment against maybe five independent macro cycles. That is a real overfitting risk, and the forward record is the only protection against it.

Glossary

Component

One of the five pieces of the regime: Liquidity, Rates, Growth, Inflation, Risk.

Reading

What a regime component says today. Five readings, not three. Full green or red means ±0.45 is crossed and the name is the full one (Hot, Tight, Strong). A lean paints the needle dim. White is only Mid / Neutral.

Needle

The score on the scale. On an asset class, lean within the call — it cannot flip the call.

1w · 2w · 1m · 3m · 6m

Lookback for the table and charts. Table default 3m. Chevrons use 1m.

In / mixed / out of favor

The asset-class call. In is green, out is red. Mixed is the reading, the needle, and the tap.

Duration risk

Whether owning long bonds is likely to hurt.

Credit risk

Whether borrowers are likely to keep paying.

Last time

A past day whose five-component mix resembles today’s. The base rate is what happened next. On a split class: one verdict, then rose-in % per sleeve. A loose match is context, not a second essay.

Archive

Today’s rules replayed over 2003 to now, rebuilt each morning. A backtest. Change a rule and all 23 years change with it. Every base rate comes from here.

Forward record

What the app actually called each morning, written down in advance and never recomputed. Began September 2026. The only number here that is a claim about the app rather than about the rules.

Independent window

Days are not tries. Hold a call for two months and grade it on the next month’s return and you have about two separate bets, not fifty — every window overlaps its neighbours. Counting the non-overlapping ones stops one long episode looking like a pile of evidence.