Global Capital Flow Intelligence System

Financial markets are driven by continuous capital movement between countries, sectors, factors, asset classes, and risk regimes. Institutional investors constantly reallocate in response to growth, inflation, monetary policy, valuations, and sentiment — yet direct fund-flow datasets (EPFR, Bloomberg, Morningstar) are expensive and inaccessible to most independent researchers.

This project constructs a Global Capital Flow Intelligence System using only Yahoo Finance ETF data. The central hypothesis is that capital flows leave measurable footprints in relative performance, volume participation, volatility structure, correlation topology, and leadership persistence. The pipeline transforms adjusted prices and volumes into a composite Capital Flow Score $F_{i,t}$, ranks assets within eight analytical layers, classifies macro regimes, builds correlation networks, and backtests flow-informed portfolios against static benchmarks.

The framework is descriptive and hypothesis-driven: it does not claim to observe actual fund flows, but infers allocation *preferences* from market-generated information across 60+ liquid ETF proxies spanning 2010–present.

The core question is: where is capital moving? Markets are driven by relative attractiveness, not absolute returns alone. Capital rotates between countries, developed and emerging markets, growth and value, equities and bonds, risk assets and defensives, sectors, currencies, and inflation-sensitive assets.

This study asks whether those rotations can be inferred when only public prices and volumes are available. Four research objectives structure the empirical work:

• Geographic allocation — US, Europe, Japan, China, India, EM, frontier proxies.
• Cross-asset allocation — equities, government bonds, credit, commodities, currencies.
• Sector and factor rotation — GICS sectors and style ETFs (momentum, value, quality, low vol, growth, dividend).
• Regime-conditional behaviour — whether flow signals and leadership differ across risk-on, risk-off, inflationary, deflationary, recovery, and crisis states.

Sustained capital inflows tend to produce strong relative performance, persistent trends, rising participation, and improving breadth. Outflows tend to produce underperformance, volatility expansion, drawdowns, and narrowing leadership.

The system does not rely on a single indicator. Each asset $i$ at date $t$ receives a vector of engineered features ${\mathbf{x}}_{i,t}$ spanning performance, participation, stability, and persistence. These are standardised cross-sectionally and combined into a scalar flow score used for ranking, heatmaps, and portfolio construction.

Daily adjusted close prices $P_{i,t}$ and volumes $V_{i,t}$ are downloaded via yfinance for 60+ ETFs. Corporate actions are handled through Yahoo adjusted closes. The panel is aligned to a common calendar with limited forward-fill (five days) for missing quotes.

Daily simple returns:

$$r_{i,t}^{(d)}=\frac{P_{i,t}}{P_{i,t-1}}-1$$

Monthly returns from month-end prices:

$$r_{i,t}^{(m)}=\frac{P_{i,t}^{ME}}{P_{i,t-1}^{ME}}-1$$

Multi-horizon trailing returns over $k$ trading days:

$$R_{i,t}^{(k)}=\frac{P_{i,t}}{P_{i,t-k}}-1$$

Implemented horizons include $k\in \{5,20,21,63,126,252\}$ days for tactical, intermediate, and strategic flow inference. Assets require at least 36 months of history to enter the scoring universe.

Realised volatility (21-day, annualised):

$${\sigma }_{i,t}=\mathrm{std}\left(r_{i,t-20:t}^{(d)}\right)\sqrt{252}$$

Relative volume (participation proxy):

$${\mathrm{RVOL}}_{i,t}=\frac{{\bar{V}}_{i,t-4:t}}{{\bar{V}}_{i,t-60:t-5}}-1$$

Current drawdown from the running maximum:

$${\mathrm{DD}}_{i,t}=\frac{P_{i,t}}{{\max }_{\tau \le t}{P}_{i,\tau }}-1$$

Moving-average trend (persistence proxy):

$${\mathrm{MAT}}_{i,t}=\frac{{\mathrm{SMA}}_{50}(P_i{)}_t}{{\mathrm{SMA}}_{200}(P_i{)}_t}-1$$

Risk-adjusted momentum:

$${\mathrm{RAM}}_{i,t}=\frac{R_{i,t}^{(63)}}{{\sigma }_{i,t}}$$

Relative return vs benchmark (SPY):

$${\mathrm{Rel}}_{i,t}^{(m)}=r_{i,t}^{(m)}-r_{\mathrm{SPY},t}^{(m)}$$

Market breadth by group $G$:

$$B_{G,t}=\frac{1}{|G|}\sum _{i\in G}1\{r_{i,t}^{(m)}>0\}$$

For each rebalance date $t$, raw features are aggregated into four dimension scores, then cross-sectionally z-scored among eligible assets ${\mathcal{U}}_t$:

$$z_{i,t}^{(d)}=\frac{x_{i,t}^{(d)}-{\mu }_t^{(d)}}{{\sigma }_t^{(d)}}$$

Dimension inputs (matching the Python implementation):

• Performance $P$: mean of $R^{(21)},R^{(63)},{\mathrm{Rel}}^{(1m)},{\mathrm{Rel}}^{(3m)},\mathrm{RAM}$.
• Participation $A$: $\mathrm{RVOL}$.
• Stability $S$: mean of $-\sigma$, $\mathrm{DD}$, $-|{\mathrm{DD}}^{\max }|$.
• Persistence $L$: mean of $\mathrm{MAT}$ and positive-month streak / 6.

The composite Capital Flow Score is a weighted sum:

$$F_{i,t}=\sum _{d\in \{P,A,S,L\}}{w}_d\cdot z_{i,t}^{(d)},w_P=0.35,w_A=0.20,w_S=0.25,w_L=0.20$$

Assets are ranked by $F_{i,t}$. Signal thresholds:

$${\text{signal}}_{i,t}=\{\begin{array}{ll}\text{strong inflow}& F_{i,t}\ge 0.5\\ \text{strong outflow}& F_{i,t}\le -0.5\\ \text{neutral}& \text{otherwise}\end{array}$$

Capital allocation is inherently relative. For assets $a$ and $b$, the $h$-month relative strength is:

$${\mathrm{RS}}_{a,b,t}^{(h)}=\prod _{j=1}^h(1+r_{a,t-j+1}^{(m)})-\prod _{j=1}^h(1+r_{b,t-j+1}^{(m)})$$

The dashboard reports canonical pairs: equities vs bonds (SPY–TLT), growth vs utilities (QQQ–XLU), EM vs US (VWO–SPY), India vs China (INDA–FXI), credit vs Treasuries (HYG–TLT), gold vs equities (GLD–SPY), momentum vs value (MTUM–VLUE), energy vs technology (XLE–XLK), Europe vs US (FEZ–SPY), and USD vs EM FX (UUP–CEW).

Within each analytical layer, 12-month relative strength vs SPY ranks regional, sector, and factor proxies.

Each layer $\ell \in \{1,\dots ,8\}$ defines a universe ${\mathcal{U}}^{(\ell )}$ of ETF proxies. Layer-level rankings use the same composite score $F_{i,t}$ restricted to ${\mathcal{U}}^{(\ell )}$:

1. Global Equity — US (SPY, QQQ, RSP, IWM), Europe (FEZ, EWU, EWG, EWQ), Asia (EWJ, FXI, KWEB, INDA, EWY, EWT), EM/frontier (VWO, ILF, EWZ, EWW, FM). 2. Fixed Income — duration (TLT, IEF, SHY), TIPS (TIP), credit (LQD, HYG, EMB), floaters/muni (FLOT, MUB). 3. Commodity — GLD, SLV, USO, UNG, CPER, DBA, DBC, URA. 4. Currency — UUP, FXE, FXY, FXF, FXA, FXC, CEW. 5. Sector — XLK, XLF, XLV, XLI, XLE, XLU, XLB, XLRE, XLP, XLY, XLC. 6. Factor — MTUM, VLUE, QUAL, USMV, IVW, SCHD, IWM. 7. Risk Sentiment — SPY, QQQ, IWM, GLD, TLT, VIX. 8. Safe Haven — GLD, TLT, UUP, FXY, FXF.

At each month-end $t$, continuous regime scores are computed from market proxies and clipped to $[0,1]$:

$${\mathrm{RiskOn}}_t=\mathrm{clip}(0.5+2\sum _{j=0}^2{r}_{\mathrm{SPY},t-j}^{(m)}-\frac{{\mathrm{VIX}}_t-18}{40})$$

$${\mathrm{SafeHaven}}_t=\mathrm{clip}(0.5+5r_{\mathrm{GLD},t}^{(m)}+5r_{\mathrm{TLT},t}^{(m)}-3r_{\mathrm{SPY},t}^{(m)})$$

$${\mathrm{Inflation}}_t=\mathrm{clip}(0.5+8r_{\mathrm{XLE},t}^{(m)}+3r_{\mathrm{GLD},t}^{(m)}-4r_{\mathrm{TLT},t}^{(m)})$$

$${\mathrm{Recovery}}_t=\mathrm{clip}(0.5+6r_{\mathrm{IWM},t}^{(m)}+4r_{\mathrm{HYG},t}^{(m)})$$

Discrete regime labels follow a priority rule: crisis if ${\mathrm{VIX}}_t>35$ or 3M SPY return $<-8\%$; risk-off if ${\mathrm{VIX}}_t>25$ and SPY $<0$; inflationary if ${\mathrm{Inflation}}_t>0.65$ and energy return $>2\%$; deflationary if bonds rally while equities and utilities are defensive; recovery if ${\mathrm{Recovery}}_t>0.6$ and SPY 3M $>2\%$; else risk-on if ${\mathrm{RiskOn}}_t>0.55$, otherwise risk-off.

Rolling 60-day Pearson correlations ${\rho }_{i,j,t}$ define a dynamic network. An edge $(i,j)$ is retained when $|{\rho }_{i,j,t}|\ge 0.35$; strong links use $|\rho |\ge 0.5$ for degree centrality:

$${\mathrm{deg}}_{i,t}=\sum _{j\ne i}1\{|{\rho }_{i,j,t}|\ge 0.5\}$$

Average market correlation (diversification thermometer):

$${\bar{\rho }}_t=\frac{2}{N(N-1)}\sum _{i<j}{\rho }_{i,j,t}$$

Leadership analysis tests whether lagged returns of a leader asset predict follower returns, e.g. credit vs equities:

$$\mathrm{corr}({\mathrm{HYG}}_t^{(3m)}-{\mathrm{TLT}}_t^{(3m)},r_{\mathrm{SPY},t+1}^{(m)})$$

Flow-informed portfolios rebalance monthly. At date $t$, the flow rotation strategy selects the top quintile:

$${\mathcal{P}}_t=\{i\in {\mathcal{U}}_t:\mathrm{rank}(F_{i,t})\le \lceil |{\mathcal{U}}_t|/5\rceil \}$$

Portfolio return with equal weights:

$$r_{\mathcal{P},t+1}^{(m)}=\frac{1}{|{\mathcal{P}}_t|}\sum _{i\in {\mathcal{P}}_t}{r}_{i,t+1}^{(m)}$$

Analogous regional, sector, and factor rotations restrict ${\mathcal{U}}_t$ to each layer universe. Multi-layer blends the three layer rotations with equal weight. Benchmarks: equal-weight universe, static 60/40 (0.6 SPY + 0.4 TLT), and buy-and-hold SPY.

Performance metrics (monthly frequency, $N$ observations):

$$\mathrm{Sharpe}=\frac{\bar{r}-r_f}{\mathrm{std}(r)}\sqrt{12},\mathrm{MDD}=\underset{\tau }{\min }(\frac{E_{\tau }}{{\max }_{s\le \tau }{E}_s}-1)$$

where $E_t={\prod }_{s\le t}(1+r_s)$ is the equity curve.

Seven formal hypotheses are evaluated. Key test statistics:

H1 — Flow scores predict returns. Monthly Spearman rank IC between $F_{i,t}$ and $r_{i,t+1}^{(m)}$; test $H_0:\mathbb{E}[\mathrm{IC}]=0$ via one-sample $t$-test across months.

H2 — Leadership persists. Month-over-month repeat rate of argmax${}_i{F}_{i,t}$; tested vs binomial null.

H3 — Sector leadership → macro. Spearman correlation between energy sector rank and forward inflation-proxy return.

H4 — Credit leads equities. Pearson correlation between HYG–TLT 3M relative strength and next-month SPY return.

H5 — Currency → regional equity. Pearson correlation between CEW 3M strength and forward VWO return.

H6 — Safe-haven lead. Pearson correlation between safe-haven basket return and forward SPY return (expect $\rho <0$).

H7 — Flow portfolio vs 60/40. Compare annualised Sharpe of flow rotation vs traditional 60/40.

Support threshold: two-sided $p<0.10$ (research default, not production trading significance).

Flow scores are inferred proxies, not EPFR-style fund flows. ETFs imperfectly represent underlying economies; Yahoo adjusted data may differ from institutional vendors; regime rules are heuristic; hypothesis tests use overlapping monthly windows; and backtests omit transaction costs.

The framework is best read as a macro allocation research platform — identifying leadership, breadth, regime state, and network structure — rather than a standalone alpha signal. Results update with each npm run data:global-capital-flow refresh.

The Results section provides the live intelligence dashboard and a Mathematical Methodology panel with the equations above applied to current data:

• Framework coverage — maps each research objective to its dashboard exhibit.
• Mathematical methodology — rendered flow-score, regime, and portfolio equations with live weights.
• Summary cards — top flow leader, current regime, risk-on/risk-off and safe-haven scores.
• Safe haven & risk sentiment engines — dedicated flight-to-quality and aggregate sentiment leaders.
• Research insights — auto-generated findings across layers, breadth, regimes, networks, and backtests.
• Flow score decomposition — $z$-score components for the top asset.
• Global flow rankings & heatmap — composite $F_{i,t}$ across the universe.
• Market breadth — $B_{G,t}$ by region, sector, and factor.
• Eight layer panels — per-layer rankings and 12M relative strength vs SPY.
• Cross-asset relative strength — pairwise ${\mathrm{RS}}_{a,b,t}^{(h)}$.
• Regime detection — five continuous scores; click legend labels to toggle lines.
• Correlation network — ${\bar{\rho }}_t$, leadership links, centrality.
• Portfolio backtests — equity curves, Sharpe, subperiod splits.
• Hypothesis testing — all seven hypotheses with test statistics.