AlphaShift Core Macro Strategy
One rule set across US stocks, Treasuries, gold and bills, long-only or leveraged, against the S&P 500 and its leveraged funds.
| Annual return (CAGR) | Volatility | Best year | Worst year | Max drawdown | Sharpe | Sortino | |
|---|---|---|---|---|---|---|---|
| AlphaShift Core · Leverage (Sharpe) | 20.7% | 21.0% | +76.5% | -11.5% | -43% | 0.92 | 1.10 |
| AlphaShift Core · Leverage (max return) | 37.0% | 33.5% | +116.7% | -12.4% | -57% | 1.06 | 1.31 |
| AlphaShift Core · Long-only | 11.4% | 9.9% | +35.2% | -5.4% | -15% | 0.97 | 1.25 |
| 3x S&P ETF | 16.7% | 52.6% | +141.5% | -86.9% | -95% | 0.53 | 0.69 |
| 2x S&P ETF | 15.4% | 35.1% | +82.1% | -69.5% | -84% | 0.54 | 0.69 |
| S&P 500 | 11.0% | 17.5% | +36.6% | -39.1% | -54% | 0.58 | 0.73 |
| 60/40 | 8.1% | 10.2% | +22.7% | -18.9% | -31% | 0.64 | 0.81 |
Walk-forward from 2004: rules re-chosen every quarter on data available at the time, traded in ETFs with costs charged. Long-only never borrows; Leverage (Sharpe) borrows up to two times capital when its rules allow, tuned for risk-adjusted return; Leverage (max return) is tuned for total return. Against them: the S&P 500, 2x and 3x daily-reset S&P funds with realistic costs, and 60/40. Leverage (Sharpe) matches the 3x fund's ending wealth at the unlevered S&P's volatility with less than half its worst fall; Long-only beats the S&P at 60/40's volatility with under a third of its worst fall.
| Annual return (CAGR) | Volatility | Best year | Worst year | Max drawdown | Sharpe | Sortino | |
|---|---|---|---|---|---|---|---|
| AlphaShift Core · Leverage (Sharpe) | 17.7% | 17.3% | +68.6% | -19.8% | -43% | 0.83 | 1.02 |
| AlphaShift Core · Leverage (max return) | 32.5% | 31.2% | +120.4% | -42.9% | -57% | 0.95 | 1.21 |
| AlphaShift Core · Long-only | 11.8% | 9.1% | +34.8% | -3.9% | -15% | 0.90 | 1.24 |
| 3x S&P ETF | 16.9% | 50.0% | +141.5% | -86.9% | -98% | 0.51 | 0.67 |
| 2x S&P ETF | 15.7% | 33.3% | +82.1% | -69.5% | -87% | 0.51 | 0.66 |
| S&P 500 | 11.8% | 16.6% | +36.6% | -39.1% | -54% | 0.55 | 0.70 |
| 60/40 | 9.8% | 10.2% | +32.6% | -18.9% | -31% | 0.64 | 0.86 |
The longer record, from 1983. Bank credit by category begins in the early 1980s, too late for a quarterly walk-forward to start there, so this record holds out every three-year block in turn, re-chooses the rules on the rest, and trades the held-out block with them. Pseudo out of sample rather than strictly causal, which is why it sits below the walk-forward. Over four decades Leverage (Sharpe) compounds at 17.7% a year with a worst fall of -43%, against the 3x fund's 16.9% at -98%: the same return at a third of the volatility, without the near-total loss of 2008.
It reads solvency and instability, and builds the portfolio from what it finds.
Two things take prices down hard. Solvency: borrowers who can no longer pay out of their income, which happens when bank lending slows and the government's flow to the private sector turns to a drain at the same time. Instability: a boom fed by so much lending, for so long, that its price depends on new credit rather than on what the asset earns. The Core Strategy reads both from the flows themselves, the government's daily cash flow and bank credit by category, before they show in spreads or prices.
Everything on this page follows from that. When the flows are strong and no boom is running out of fuel, the strategy is fully invested, tilted toward the sectors the growth and inflation readings favour. When solvency and instability turn together, it steps aside into Treasuries, gold and bills, and the leveraged form hands its borrowing back. The backtest above is what that reading earned; the sections below show the reading on its own.
Solvency
Can borrowers keep paying? Read from the two flows that create the income debts are serviced from: bank lending and the government's net spending.
Instability
Has a boom outrun the income behind it? Read from where lending has concentrated, how long it has run, and whether the price is holding while the lending fades.
The portfolio
Fully invested while the money is flowing and the market is steady; out of the way when both readings turn. Long-only, or with borrowing that comes and goes with the credit cycle.
Where each credit-fed boom stands
The instability reading as a state machine, shown against the market since 1983.
A lending-fed boom is stable only while the lending continues; Minsky's point was that the stability itself breeds the fragility. The instability clock tracks each channel of credit, housing, corporate lending, duration and non-bank lending, through three acts: a boom forming, the point where it turns fragile, and the window in which it breaks. An aftermath follows each pop until the cycle clears. What moves it from one act to the next stays with us.
Every week since 1983, shaded by the act the instability clock was in, over the S&P 500. The pop windows sit at or near the tops of 1987, 2000, 2007 and 2022. Fast shocks with no credit boom behind them, 1998, 2018 and 2020 among them, it cannot see, which is why the strategy carries a diversified book at all times. Forming booms pay, so the strategy stays invested; in the later acts it steps aside, and in the pop window it is out.
The theory holds in other macro landscapes
The same rules, untouched, run on four other economies with their own data.
The rules were written on US data, and the fair objection is that four US crises could have been learned by heart. So the whole machine was run unchanged on the United Kingdom, Canada, Australia and Japan: their own fiscal, credit and market data, none of it looked at when the rules were written, and each with a credit-fed bust around 1990 that the US did not have. A different universe, the same theory.
The instability clock saw those booms: in each chart the late-1980s boom arms and moves through its acts ahead of the country's own bust. The strategy went defensive in every one, and again in 2008, beating local 60/40 and local equities in each bust year. The edge is smaller than in the US, as rules tuned on one country should show; the mechanism is what transfers.
What the strategy did in each country's bust, and in 2008
Calendar-year returns of the transferred long-only strategy against local 60/40 and local equities, in each country's bust year and in 2008, with the Sharpe ratio over the whole window.
| Country | Window | The bust | Bust year: strategy / 60/40 / equities | 2008: strategy / 60/40 / equities | Sharpe: strategy / 60/40 |
|---|---|---|---|---|---|
| United Kingdom | 1986–2026 | 1989–90 housing bust | +2% / -4% / -10% | -1% / -18% / -34% | 0.37 / 0.36 |
| Canada | 1980–2026 | 1990–92 housing and credit bust | -5% / -9% / -16% | -9% / -21% / -38% | 0.31 / 0.41 |
| Australia | 1986–2026 | 1990–91 recession and property bust | +11% / -4% / -18% | +2% / -24% / -43% | 0.41 / 0.39 |
| Japan | 1976–2026 | 1990 bubble pop | -17% / -26% / -38% | -12% / -26% / -42% | 0.59 / 0.43 |
What the stock signal alone is worth
The strategy's stock timing separated from everything else it does, at the S&P's volatility.
What this isolates. The stock decision alone. A mechanical trader reads only the strategy's stock probabilities every four weeks and holds a long S&P position sized by a fixed rule (Kelly on the lower bound of the expected 13-week return, 12% volatility target, never above one times capital, never short), the rest in bills. No bonds, no gold, so it earns nothing while out of stocks.
How to read it. Growth of $100,000 from 2004, log scale, every line scaled to the S&P's volatility (17.5%), so a higher ending value is higher risk-adjusted return, not more exposure. The multiplier in the legend is the leverage the scaling implied. A scaling of returns, not a traded line.
What it shows. At the market's volatility the stock signal alone ends at $1,799,018 against the S&P's $1,066,920, with a worst fall of -41% against -54%: there is information in the stock decision on its own. The long-only Core Strategy, scaled the same way, ends at $4,987,916. The gap between the two is what the rest of the strategy adds: bonds and gold held alongside stocks, and moved into when the stock signal steps aside.
| Leverage applied | Annual return | Volatility | Max drawdown | Sharpe | $100,000 becomes | |
|---|---|---|---|---|---|---|
| AlphaShift Core · Long-only | 1.78× | 18.7% | 17.5% | -26.5% | 0.97 | $4,987,916 |
| Stock signal alone, traded by formula | 1.73× | 13.5% | 17.5% | -40.6% | 0.71 | $1,799,018 |
| S&P 500 | 1.00× | 11.0% | 17.5% | -54.0% | 0.58 | $1,066,920 |
| 60/40 | 1.72× | 12.1% | 17.5% | -49.6% | 0.64 | $1,341,159 |
The probability of a fall, conditional on the strategy's positioning
A portfolio return mixes the quality of a prediction with the noise of the assets it was applied to. This section removes the portfolio.
How the probabilities are built. Each week the Core Strategy's equity share is placed in one of five bins: under 10%, 10 to 25%, 25 to 45%, 45 to 85%, 85% and above. For each bin, window (4, 13, 26, 52 weeks) and move size (5, 10, 20, 30%), the probability is the share of weeks in that bin followed by that move inside the window, measured from the running high. The live feed computes these walk-forward; the funnel pools the full history to show the shape.
How to read a panel. Rows are move sizes, columns are windows. The five bars are the equity-share bins, lightest on the left, fully invested on the right. Bar height is the probability, the whisker a 90% range from a block bootstrap, n the number of distinct episodes. Grey bars rest on fewer than five episodes.
What the fall funnel shows. In every panel with enough episodes the bars step down from left to right: falls are most likely when the strategy holds the least stock, least likely when it is fully invested. For a 10% fall inside 13 weeks the probability runs from 35% in the lightest bin to 4% at full stock. Today, fully invested, the 13-week probability of a 10% fall is 6% against an unconditional 13%, and of a 20% fall 0% against 2%. Out of sample from 1990, the walk-forward probability ranks 10% falls inside a quarter at an AUC of 0.64.
The surface today (2026-09-04: the strategy 100% in stocks, instability clock forming, VIX 14)
What this is. The same construction live, for this week, at the full grid of windows and sizes, in four forms: falls and rises each as a path event (a move of that size from the running high or low at any point inside the window) and as an end-point event (the return over the whole window beyond that size). Walk-forward: each cell uses only the weeks that shared today's positioning and instability-clock state, with outcomes known by that date.
How to read a cell. The probability, then the unconditional rate in parentheses, the 90% range in brackets, and n distinct episodes. Red is at least a quarter above the unconditional rate, blue at least a fifth below; grey rests on fewer than five episodes. The last table is the end-point distribution as quantiles.
| size | 4 weeks | 8 weeks | 13 weeks | 26 weeks | 52 weeks |
|---|---|---|---|---|---|
| 2% | 30% (36%) [21%–40%] n27 | 63% (63%) [48%–75%] n29 | 78% (82%) [60%–90%] n22 | 95% (97%) [70%–99%] n25 | 100% (100%) [62%–100%] n20 |
| 3% | 22% (23%) [15%–32%] n22 | 52% (48%) [37%–65%] n26 | 72% (68%) [53%–85%] n23 | 93% (91%) [68%–99%] n25 | 100% (99%) [62%–100%] n20 |
| 5% | 7% (10%) [4%–15%] n8 | 22% (25%) [12%–36%] n15 | 43% (40%) [26%–61%] n16 | 66% (64%) [39%–85%] n20 | 97% (86%) [59%–100%] n21 |
| 10% | 1% (2%) [0%–5%] n1 | 4% (7%) [1%–14%] n1 | 6% (13%) [1%–21%] n1 | 14% (27%) [4%–41%] n3 | 43% (47%) [15%–76%] n10 |
| 15% | 0% (0%) [0%–4%] n0 | 1% (3%) [0%–10%] n1 | 3% (6%) [1%–17%] n1 | 9% (16%) [2%–35%] n1 | 27% (33%) [7%–64%] n5 |
| 20% | 0% (0%) [0%–4%] n0 | 0% (1%) [0%–8%] n0 | 0% (2%) [0%–12%] n0 | 0% (6%) [0%–23%] n0 | 7% (13%) [1%–47%] n2 |
| 25% | 0% (0%) [0%–4%] n0 | 0% (1%) [0%–8%] n0 | 0% (1%) [0%–12%] n0 | 0% (4%) [0%–23%] n0 | 7% (11%) [1%–47%] n2 |
| size | 4 weeks | 8 weeks | 13 weeks | 26 weeks | 52 weeks |
|---|---|---|---|---|---|
| 2% | 16% (20%) [10%–24%] n21 | 17% (20%) [8%–30%] n15 | 17% (21%) [7%–35%] n13 | 11% (20%) [2%–37%] n7 | 10% (17%) [1%–49%] n3 |
| 3% | 10% (15%) [6%–18%] n14 | 14% (17%) [7%–27%] n14 | 14% (18%) [5%–31%] n11 | 9% (17%) [2%–35%] n6 | 10% (16%) [1%–49%] n3 |
| 5% | 5% (8%) [2%–12%] n6 | 6% (11%) [2%–17%] n7 | 7% (13%) [2%–23%] n6 | 6% (13%) [1%–31%] n4 | 9% (15%) [1%–48%] n3 |
| 10% | 1% (2%) [0%–5%] n2 | 2% (3%) [0%–11%] n2 | 2% (6%) [0%–16%] n2 | 2% (7%) [0%–25%] n2 | 1% (10%) [0%–40%] n3 |
| 15% | 0% (0%) [0%–4%] n0 | 1% (1%) [0%–9%] n1 | 0% (2%) [0%–13%] n1 | 0% (4%) [0%–23%] n1 | 0% (6%) [0%–38%] n0 |
| 20% | 0% (0%) [0%–4%] n0 | 0% (1%) [0%–8%] n0 | 0% (1%) [0%–12%] n0 | 0% (2%) [0%–23%] n0 | 0% (3%) [0%–38%] n0 |
| 25% | 0% (0%) [0%–4%] n0 | 0% (0%) [0%–8%] n0 | 0% (1%) [0%–12%] n0 | 0% (1%) [0%–23%] n0 | 0% (2%) [0%–38%] n0 |
| size | 4 weeks | 8 weeks | 13 weeks | 26 weeks | 52 weeks |
|---|---|---|---|---|---|
| 2% | 59% (58%) [49%–68%] n38 | 86% (91%) [73%–93%] n27 | 97% (99%) [83%–99%] n24 | 100% (100%) [77%–100%] n22 | 100% (100%) [62%–100%] n20 |
| 3% | 43% (37%) [34%–54%] n34 | 75% (78%) [61%–85%] n25 | 93% (95%) [77%–98%] n26 | 100% (100%) [77%–100%] n22 | 100% (100%) [62%–100%] n20 |
| 5% | 14% (13%) [9%–23%] n17 | 54% (49%) [39%–67%] n23 | 78% (75%) [60%–90%] n21 | 97% (97%) [73%–100%] n21 | 100% (100%) [62%–100%] n20 |
| 10% | 1% (1%) [0%–5%] n2 | 13% (10%) [6%–25%] n9 | 32% (30%) [18%–51%] n14 | 79% (72%) [52%–93%] n20 | 95% (97%) [56%–100%] n20 |
| 15% | 0% (0%) [0%–4%] n0 | 4% (2%) [1%–15%] n5 | 19% (9%) [8%–37%] n9 | 55% (42%) [30%–78%] n19 | 87% (83%) [48%–98%] n19 |
| 20% | 0% (0%) [0%–4%] n0 | 1% (1%) [0%–9%] n1 | 6% (3%) [2%–22%] n5 | 35% (20%) [15%–61%] n10 | 79% (63%) [41%–95%] n16 |
| 25% | 0% (0%) [0%–4%] n0 | 0% (0%) [0%–8%] n0 | 1% (1%) [0%–14%] n1 | 18% (9%) [5%–45%] n5 | 71% (48%) [34%–92%] n15 |
| size | 4 weeks | 8 weeks | 13 weeks | 26 weeks | 52 weeks |
|---|---|---|---|---|---|
| 2% | 49% (43%) [39%–59%] n38 | 55% (54%) [40%–68%] n28 | 63% (61%) [45%–79%] n27 | 81% (71%) [54%–94%] n21 | 86% (79%) [46%–98%] n19 |
| 3% | 37% (31%) [28%–47%] n33 | 48% (47%) [34%–62%] n32 | 61% (56%) [42%–77%] n23 | 78% (68%) [51%–92%] n21 | 84% (77%) [45%–97%] n19 |
| 5% | 18% (14%) [11%–27%] n24 | 34% (31%) [22%–48%] n21 | 51% (43%) [33%–68%] n23 | 74% (61%) [46%–90%] n23 | 83% (74%) [44%–97%] n17 |
| 10% | 2% (2%) [0%–7%] n4 | 13% (7%) [6%–25%] n12 | 23% (16%) [11%–42%] n14 | 55% (37%) [30%–78%] n17 | 75% (65%) [37%–94%] n21 |
| 15% | 0% (0%) [0%–5%] n1 | 3% (2%) [1%–13%] n5 | 14% (5%) [5%–31%] n9 | 35% (20%) [16%–62%] n21 | 66% (51%) [30%–90%] n16 |
| 20% | 0% (0%) [0%–4%] n0 | 0% (0%) [0%–9%] n1 | 4% (2%) [1%–18%] n5 | 19% (10%) [6%–46%] n11 | 47% (35%) [17%–79%] n19 |
| 25% | 0% (0%) [0%–4%] n0 | 0% (0%) [0%–8%] n0 | 0% (0%) [0%–13%] n1 | 8% (4%) [1%–33%] n7 | 32% (22%) [9%–69%] n15 |
| Window | 5th | 10th | 25th | Median | 75th | 90th | 95th | Weeks in the set · effective |
|---|---|---|---|---|---|---|---|---|
| 4 weeks | -5.3 | -3.1 | -0.8 | +1.9 | +4.1 | +6.1 | +8.1 | 258 · 64 |
| 8 weeks | -5.2 | -4.2 | -0.5 | +2.5 | +6.5 | +11.3 | +13.3 | 254 · 32 |
| 13 weeks | -7.1 | -3.9 | -0.0 | +5.2 | +9.6 | +16.2 | +17.5 | 251 · 19 |
| 26 weeks | -5.5 | -2.5 | +4.3 | +11.7 | +17.0 | +23.5 | +27.4 | 238 · 9 |
| 52 weeks | -7.5 | -1.5 | +9.8 | +19.2 | +29.1 | +36.2 | +37.6 | 229 · 4 |
Bets against random timing
The over- and under-weights scored as bets, with the return of simply owning the assets stripped out, against the same bets at random dates.
How the bets are defined. For each of the three sleeves, stocks, bonds and bills, gold and oil, the long-run average holding is the passive position and is removed. What is left each week is the bet: the amount above or below that average, scored on the sleeve's excess return over bills the following week. The result is the profit from timing alone. The strategy never sells short; an underweight is a short against its own average.
How to read the chart. The black line is cumulative bet profit as a fraction of capital. Each grey line is one of 300 placebo runs with the same bets moved to random dates in half-year blocks; the band is their 10th to 90th percentile. With no information the black line would sit inside the band. The steps up in 1987, 2000 to 2002, 2008, 2020 and 2022 are underweights during large falls.
What it shows. Stock bets earn +2.1% a year and beat 100% of the 300 random timings. All three sleeves together earn +3.1% a year at a Sharpe of 0.52, against a random median of -0.00 and a random 90th percentile of 0.20.
Every fall of 15% or worse since 1983
One row per S&P fall of 15% or more, peak to trough: the S&P in red, 60/40 in grey and the long-only Core Strategy in blue over the same dates, with the strategy's own worst fall inside the episode on the right. It held under half its capital in stocks at the peak in seven of the nine episodes; the two it did not anticipate, 1998 and 2025, were fast shocks, where diversification held its loss to about half the market's.
DeepMinsky: a working model of the economy, run forward from today
The readings say where the cycle is. DeepMinsky is how we ask where it is going: a system-dynamics model of the whole economy, households, firms, banks, government and the rest of the world, each with a balance sheet, linked by the accounting that makes one sector's spending another's income. Banks create money when they lend, the government when it runs a deficit; profits follow Kalecki, credit builds and breaks the way Minsky described. Nothing is fitted to history.
Each week the model takes the latest data and runs forward: a path with a range for growth, inflation, unemployment, rates, and the two flows that matter most to the book, credit and fiscal creation. The range widens the further out it looks.
How the book uses it
The strategy is read twice: on today's readings, and on where the model says they are going. Agreement is confirmation; a forecast leaning against the data is the early warning.
What it is good for
The direction of growth, inflation and the two flows over the next two to four quarters, the shape of a recession once credit turns, and what a fiscal or credit shock does to the rest of the system.
What it is not
A price forecast. It says nothing about the S&P next month, only where the income and money that support prices are headed.