> ## Content Index
> Fetch the complete content index at: https://aligrithm.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# 4.78 Gold vs TIPS: Competing Safe Assets and the Post-2022 Decoupling
- URL: https://aligrithm.com/gold-vs-tips-competing-safe-assets-and-the-post-2022-decoupling/
- Published: 2026-10-28T20:01:00.000Z
- Updated: 2026-10-28T20:00:59.000Z
- Description: Gold versus 10-year TIPS: a 100 bp real-yield rise lined up with a 6.4% drop in log gold. After 2022 that weekly slope is -0.01 (t = -0.98) and gold still rose 0.50% a week.
- Author: ali askar
- Tags: 4. Market Structure Notes

Batten, Loncarski, Szilagyi, and Zhou put a long gold book next to the 10-year real yield and the dollar. On monthly data from February 2003 to February 2026, the log change in the London gold fix falls by 0.0639 for each percentage point the 10-year TIPS yield rises, holding the ICE dollar index fixed. A 100 basis point backup, dollar unchanged, is a 6.39% drop in the log gold price. That slope is the opportunity-cost hedge a gold book usually carries: the metal pays no coupon, so a higher real yield is income you give up by holding it. After 25 March 2022 the weekly version of the same slope is -0.0125, with a t-statistic of -0.98, and the fitted intercept is still +0.50% a week.

## A 100 basis point rise lines up with a 6.39% drop in log gold

The monthly bivariate fit, 278 months, puts the TIPS slope at -0.0639 (t = -4.17) and the dollar slope at -0.7469 (t = -4.64), with an R-squared of 0.290\. TIPS is the 10-year Treasury Inflation-Protected Security. You compute its yield as a real yield because the principal steps with the consumer price index, so the coupon is already in purchasing-power terms. The left side is the change in the natural log of the London gold fix. R-squared is the share of that log-return variance the two regressors line up with.

$$ \\Delta \\ln S\_t = \\alpha + \\beta\_{\\text{TIPS}} \\, \\Delta y\_{10,t}^{\\text{TIPS}} + \\beta\_{\\text{DXY}} \\, \\Delta \\ln \\text{DXY}\_t + \\varepsilon\_t $$

The change in log gold equals an intercept, plus a slope on the change in the 10-year TIPS yield, plus a slope on the log change in the ICE US Dollar Index, plus a residual. Units matter, and the paper uses two conventions for the TIPS change. Table 1 reports gold and dollar returns in percent and the TIPS change in basis points. The regression rescales: divide a percent return by 100 to get the decimal log change, and divide a basis-point yield change by 100 to get percentage points. A TIPS change of 1.00 means 100 basis points. A dollar change of 0.01 means a 1% rise in the index. One equation line calls the TIPS yield "decimal form." Their own worked product uses percentage points. A true decimal yield, 0.01 for 100 basis points, times -0.0639 would move log gold by -0.000639, which is -0.06%, and that is not the -1.45% they report for a one-standard-deviation yield move. Keep the percentage-point reading.

Work the 100 basis point case with the dollar unchanged. The TIPS term is -0.0639 times 1.00, so the log gold price falls by 0.0639, which is -6.39%. The price ratio is the exponential of that log change: exp(-0.0639) is 0.938, so the fix is 6.19% lower. At this size the log and the simple return differ by 0.20 percentage points. Use the log number when you are adding months, and the simple number when you are marking a book to a price.

A one-standard-deviation month is smaller and it checks. Table 1 puts the monthly TIPS change at a standard deviation of 22.694 basis points, which is 0.22694 percentage points. Times -0.0639 that is -0.01450, or -1.45% in the log price. Monthly gold volatility in the same table is 4.796%, and 1.45 / 4.80 is 30%. Their sentence matches. The dollar's monthly standard deviation is 2.215%, or 0.02215 in the regression's units. Times -0.7469 that is -0.01654, or -1.65%. Against the same 4.80% gold volatility that is 34%. Their prose calls the same ratio close to 30%. The -1.66% they print is the product with the rounded slope -0.747\. The 30% label on the dollar leg is loose.

They also apply the rounded slopes to the 2022 to 2023 tightening: TIPS yields up about 300 basis points, the dollar up about 15%. The printed product is (-0.064 times 3.00) plus (-0.747 times 0.15) = -0.305, "roughly -30.5% on a compound basis." With the table slopes, -0.0639 times 3.00 plus -0.7469 times 0.15 equals -0.304\. Either way the sum is a log return, the left side of the regression. The compounded price change is exp(-0.304) - 1, which is -26.2%. They labeled a log return of -0.30 as a -30.5% compound price change. Keep -0.30 as the fitted log move. Mark the book at about -26%.

Translate the same slope into a modified duration. For a small yield move the percentage price change is minus duration times the yield change, and the yield change in that definition is in decimal units, 0.01 per 100 basis points.

$$ \\frac{\\Delta S}{S} \\approx -D \\, \\Delta y\_{\\text{decimal}}, \\qquad D = \\frac{-\\beta\_{\\text{TIPS}}}{0.01} $$

Duration is the log-price response per percentage point of yield, divided by 0.01, because 100 basis points is 0.01 in decimal. On the monthly slope, D is 0.0639 / 0.01 = 6.39 years. A gold book that loads at -0.0639 behaves, on this sample, like a zero-coupon claim with about 6.4 years of real-yield duration. The appendix prints Duration = minus beta-TIPS over the real-rate change = 0.073, using the zero-rate slope below, and then calls 0.073 "an effective gold duration of approximately 7.3 years." The ratio 0.073 is the decimal price change for a one-percentage-point move. The duration in years is that price change divided by 0.01\. The 7.3-year sentence is the right economic reading. The printed rearrangement drops the division. The source may have moved the decimal in prose and left the formula behind.

The variance split in the prose does not match Table 3 either. TIPS alone has an R-squared of 0.189, which is the square of the monthly gold-TIPS correlation of -0.435\. The dollar alone has an R-squared of 0.213, the square of -0.461\. Together they reach 0.290\. The prose says the yield channel accounts for about 13% and the dollar for about 14%. Those shares are not the univariate R-squareds, and they are not the increments either: adding TIPS to the dollar raises R-squared by 0.290 minus 0.213, which is 0.077\. Use 0.19, 0.21, and 0.29\. Seventy-one percent of monthly gold variance is still the residual.

Break-even inflation does not take any of that residual. Break-even is the 10-year nominal Treasury yield minus the 10-year TIPS yield, the inflation compensation priced into the nominal bond. Its slope, once TIPS and the dollar are in the regression, is +0.009 with a t-statistic of 0.58, and R-squared moves from 0.290 to 0.291\. A 100 basis point rise in break-even, real yield and dollar held fixed, is a 0.9% log-gold move you cannot separate from zero. The old article ["Why FX Traders Must Watch Gold, Rates, and Equities"](https://aligrithm.com/why-fx-traders-must-watch-gold-rates-and-equities/) is right that a gold book has to watch rates. This regression says which rate. Watch the real yield. Break-even's slope is the one that failed. A change in the VIX, the CBOE equity-volatility index, enters at +0.0011 per index point (t = 1.97) and lifts R-squared from 0.290 to 0.302\. Five VIX points is 0.55% of log gold. Small next to either of the two slopes you already have.

The weekly fit, January 2006 to February 2026, 1,052 weeks, is the same equation at a faster clock. The TIPS slope is -0.0325 (t = -2.71), about half the monthly slope, and the dollar slope is -0.953 (t = -10.11), about 28% larger than -0.747, with an R-squared of 0.241\. The dollar is in the gold fix the same week. The real-yield adjustment is slower, which is why giving it a month roughly doubles its slope. The t-statistics here and above use Newey-West standard errors, 4 lags on the monthly sample and 8 on the weekly, so they allow residuals that bunch and that correlate with their own past. The bandwidth formula they print, the floor of 4 times (T/100) to the power 2/9, equals 5 for 278 months and 6 for 1,052 weeks. They used 4 and 8\. The slopes are ordinary least squares, so the lag does not move them. Table 18 keeps the monthly slopes at -0.0639 and -0.7469 across the lags they tried. The lag choice moves the t-statistics.

## The monthly slope is steeper below a TIPS yield of zero

Below a 10-year TIPS yield of zero the monthly slope is -0.0912 (t = -3.70), on 49 months, with an R-squared of 0.385\. Above zero it is -0.0625 (t = -3.77), on 229 months, with an R-squared of 0.283\. The below-zero slope is larger in magnitude by (0.0912 - 0.0625) / 0.0625, which is 46%. They call the rounded pair, -0.091 and -0.063, "roughly 45% stronger." That rounding checks. The larger slope sits on 49 months.

$$ \\Delta \\ln S\_t = \\alpha + \\beta(y\_t^{\\text{TIPS}}) \\, \\Delta y\_{10,t}^{\\text{TIPS}} + \\beta\_{\\text{DXY}} \\, \\Delta \\ln \\text{DXY}\_t + \\varepsilon\_t $$

The TIPS slope depends on the yield level you are standing on when the yield changes. They implement that by splitting the sample. The Hansen threshold test asks whether a kink in an observable, here the TIPS yield level, fits better than one slope, and it estimates the kink rather than taking zero as given. On weekly data the likelihood-ratio statistic is 53.3, bootstrap p-value below 0.001, and the estimated kink is 1.78%. On monthly data the statistic is 16.4, the kink is 0.61%, and the p-value is 0.114\. The monthly 10% critical value in their table is 16.99, so 16.4 does not reject a single monthly slope.

Two gaps sit between that test and a trading switch. They never print the regression split at 1.78%. The coefficients you can recompute are the split at zero. And the weekly split at zero does not repeat the monthly pattern cleanly: below zero the weekly slope is -0.0954 with a t-statistic of -1.58 on 205 weeks, which does not clear 5%, and above zero it is -0.0199 with a t-statistic of -7.07 on 847 weeks. That precise -0.020 is a blended weekly average of pre-crisis weeks, zero-rate weeks that happened to be slightly positive, and the post-2022 weeks.

I am against using 1.78% as a hard switch on a monthly book. The monthly test missed, and the two slopes at that kink are not in the paper. What you can compute is the coarser monthly split at zero, on a thin negative-yield sample of 49 months, and the weekly time-breaks in the next section.

## Inside the zero-rate years the TIPS slope builds for half a year

From January 2009 to March 2022 the one-week TIPS slope is -0.073 and the 26-week slope is -0.188\. The dollar slope over those same windows stays between -0.67 and -0.82\. The currency move is in the gold price at the one-week horizon. The gold response to the yield change is still larger when both are measured over 26 weeks.

$$ \\Delta \\ln S\_t^{(h)} = \\alpha\_h + \\beta\_{\\text{TIPS}}(h) \\, \\Delta y\_t^{(h)} + \\beta\_{\\text{DXY}}(h) \\, \\Delta \\ln \\text{DXY}\_t^{(h)} + \\varepsilon\_t^{(h)} $$

Same regression, now on non-overlapping h-week changes. A horizon of 26 means you sum 26 weeks of log gold, 26 weeks of TIPS changes, and 26 weeks of dollar changes, then roll forward by 26 weeks so the windows do not overlap. The slope is still a response per percentage point of the h-week yield change. If slow buyers finish a rebalance over months, a 100 basis point move measured over half a year shows a larger gold response than the same 100 basis points crammed into one week, because the one-week window cuts them off.

Look at the blue points. They step from about -0.07 at one week to about -0.19 at 26 weeks. The grey post-2022 points stay near zero. The red pre-crisis points start positive, and the error bars on the longer red estimates swallow the point.

![TIPS and dollar slopes by horizon, non-overlapping windows, for the pre-crisis, zero-rate, and post-2022 samples](https://storage.ghost.io/c/27/cb/27cb0fc8-2c77-4434-af9e-d5d32a916994/content/images/2026/10/article_472-tips_beta_by_horizon.png)

Work the 26-week zero-rate slope. A 100 basis point rise over those 26 weeks, dollar unchanged, is -0.188 in log gold, and exp(-0.188) - 1 is a price change of -17.1%. The R-squared on that row is 0.82\. The sample is 26 non-overlapping windows, and the t-statistic is -10.12\. A t-statistic that large on 26 points is what a tight in-sample fit looks like when you have two slopes and almost no leftover variance. Use the direction, from -0.073 at one week to -0.188 at 26 weeks. Do not paste -0.188 into a live hedge ratio.

Table 5 and Table 12 do not agree on the standard error of the one-week zero-rate slope. Both print a coefficient of -0.073, an R-squared of 0.31, and 688 weeks. Table 5 reports a t-statistic of -7.63\. Table 12 reports -8.51\. Keep -0.073\. The source disagrees with itself on the precision.

After March 2022 the one-week slope in Table 12 is -0.013 (t = -1.00). The eight-week slope is -0.044 (t = -2.22) on 25 windows, and the 13-week slope fades to -0.032 (t = -1.79) on 15 windows. That eight-week cell is the only post-2022 horizon that clears 5%.

## After March 2022 the real-yield bar crosses zero

Look at the grey bar in panel (a). It sits on zero. The red bar is positive and the blue bar is the substitution slope.

![TIPS slope, dollar slope, and R-squared in the pre-crisis, zero-rate, and post-2022 weekly regimes](https://storage.ghost.io/c/27/cb/27cb0fc8-2c77-4434-af9e-d5d32a916994/content/images/2026/10/article_472-regime_coefficients.png)

Three numbers, from the weekly bivariate fits, rounded to two decimals: the TIPS slope is +0.04 before the crisis, -0.07 from January 2009 to March 2022, and -0.01 after 25 March 2022\. The rest of that grid: dollar slopes -1.70, -0.75, and -0.94, R-squared 0.33, 0.31, and 0.19, on 158, 688, and 206 weeks. The t-statistics on the TIPS slope are +2.40, -7.63, and -0.98\. Bai-Perron, the break test that searches for coefficient shifts without you naming the dates, places the breaks on 16 January 2009 and 25 March 2022.

They report results from about 70 specifications. A Bonferroni cut that divides 5% by 70 is a critical absolute t-statistic of about 3.39\. The zero-rate t-statistic of -7.63 still clears it. The post-2022 t-statistic of -0.98 does not, and no multiple-testing adjustment creates a slope that is already indistinguishable from zero.

The monthly break test does not find these dates. Its sup-F is 5.23 against a 5% critical value of 11.47, so a single monthly slope is not rejected. Rolling 60-month windows keep a negative TIPS slope, somewhere between about -0.002 and -0.164, with a mean of -0.091\. That roll cannot yet be a clean post-2022 estimate. March 2022 to February 2026 is about four years, short of 60 months, so every 60-month window that reaches the end of the sample still contains zero-rate years. If you check the monthly roll and "still see a negative beta," you are looking at a window that has not dropped 2009 to 2022.

Look at the shaded window to the right of the March 2022 line. The TIPS yield is positive and the gold price rises through it.

![London gold fix against the 10-year TIPS yield, with the January 2009 and March 2022 breaks](https://storage.ghost.io/c/27/cb/27cb0fc8-2c77-4434-af9e-d5d32a916994/content/images/2026/10/article_472-gold_and_tips_yields.png)

Their round figures for that window: the 10-year TIPS yield went from about -1% to about +2.5%, and gold went from around $2,050 in March 2022 to more than $2,900 by February 2026\. Apply the zero-rate slope to the yield piece alone, dollar unchanged.

$$ \\Delta \\ln S \\approx (-0.0729) \\times 3.50 = -0.255 $$

A 3.50 percentage point rise in the real yield, times the zero-rate slope of -0.0729, is a log gold change of -0.255\. The price change is exp(-0.255) - 1, or -22.5%. The same round prices give 2900 / 2050 - 1 = +41.5%, and the log change is +0.347\. The old yield channel, on the yield move alone, called for a decline of about a quarter. The fix rose by about two-fifths. The dollar also moved over those four years, so this arithmetic is the yield term alone. A full residual would include the dollar's path over the same four years, which they do not print as a single number. The same division that turned -0.0639 into 6.39 years of duration turns -0.0729 into 7.29 years and turns -0.0125 into 1.25 years. The real-yield duration you would have been hedging fell from about 7.3 years to about 1.3.

The fitted intercept in that third regime is +0.0050 per week, which is +0.50% in the log price in a week when the TIPS yield and the dollar do not change. Multiply by 52 and you get +26%, the annualization they print. Compound the same weekly rate, (1.005) to the 52nd power minus 1, and you get +29.6%. They printed the linear one. Treat +26% as 52 times the weekly intercept. The compounded price return on that same weekly rate is the +29.6%.

Table 11's post-2024 row goes further than "insignificant." The TIPS slope is +0.0035 with a standard error of 0.0202, on 113 weeks, R-squared 0.122\. The point estimate has changed sign. Section 6 of the paper quotes subperiod slopes of -0.067, -0.057, and -0.059\. Those numbers are not in Table 11\. The table's Post-GFC/ZIRP, low-rate, and COVID rows print -0.049, -0.090, and -0.119\. Use the table. The narrative coefficients and the printed subperiod table do not describe the same splits, and the source may have mixed windows.

A three-state Markov-switching regression, which lets the slope and the residual variance jump with an estimated probability rather than on a fixed date, finds a state with a TIPS slope of -0.080 (t = -9.49) that sits on the zero-rate years and matches the -0.073 break estimate within about 10%. The other states lump the pre-crisis weeks with the post-2022 weeks, because both have a weak TIPS channel and a fatter residual. The specialist-holding label and the post-2022 buyer are yours. The likelihood only sees the shared weak slope. The old article ["Stock-Bond Correlation: The Sign Flips and So Do Its Drivers"](https://aligrithm.com/stock-bond-correlation-the-sign-flips-and-so-do-its-drivers/) is the same kind of object on the stock-bond pair: the sign of the safe-asset link is a real-rate regime, and a model that only sees coefficients will glue together episodes that only share a dead loading.

With contemporaneous regressors, an expanding window that starts on half the sample and steps forward one observation, the out-of-sample R-squared is 0.46 in the zero-rate regime, 0.27 before the crisis, and 0.06 after 2022\. Same-week attribution. The mean squared error of the bivariate fit is 0.712 times the historical mean on monthly data and 0.708 times it on weekly data. That is a description of the week you are in, once you already know that week's yield change and that week's dollar move.

![](https://storage.ghost.io/c/27/cb/27cb0fc8-2c77-4434-af9e-d5d32a916994/content/images/2026/10/article_472-visual-0.png)

## Size the real-yield hedge off the slope you are in

Last period's TIPS change and last period's dollar move do not forecast this period's gold return. They describe the lagged regression and do not print the lagged slopes. What they print is the out-of-sample mean squared error against a forecast that ignores both variables and uses gold's historical average return.

$$ \\Delta \\ln S\_t = \\alpha + \\beta\_{\\text{TIPS}} \\, \\Delta y\_{10,t-1}^{\\text{TIPS}} + \\beta\_{\\text{DXY}} \\, \\Delta \\ln \\text{DXY}\_{t-1} + \\varepsilon\_t $$

Gold's log return this period on last period's yield change and last period's dollar change. The source did not print those two slopes. The monthly mean-squared-error ratio versus the historical mean is 1.031, which is 3.1% worse, on 139 out-of-sample months. The Diebold-Mariano statistic, the test of whether that squared-error gap is noise, is t = 1.73 (p = 0.083). The weekly ratio is 1.007, t = 0.85, on 526 out-of-sample weeks. A ratio above 1 means the lag loses to "gold returned its sample average." The old article ["Stocks and Bonds Predict FX — But Only in Emerging Markets"](https://aligrithm.com/stocks-and-bonds-predict-fx-but-only-in-emerging-markets/) already showed a cross-asset link that is alive in one sample and gone outside it. On this book the same-week fit is real in the zero-rate years, and the lag does not forecast.

So the slope is a hedge ratio, and the hedge ratio changed. The zero-rate weekly slope is -0.0729\. The post-2022 weekly slope is -0.0125\. The gap is -0.0604 per percentage point of yield. On a 100 basis point move, that is 6.04% of log notional you are still hedging if you kept the old ratio. On a $10 million gold book, 0.0604 times 10 million is $604,000 of log exposure against a yield move that, in the later window, has a fitted gold response of 1.25%. The zero-rate slope would have put that same 100 basis point move at 7.29%. The simple price gap on that 6.04% log difference is exp(-0.0604) - 1, or -5.86%, which is $586,000 on the same book. There is no turnover in the paper, so there is no commission to subtract. The cost it left out is that hedge error.

The dollar hedge is still in the later window. The post-2022 dollar slope is -0.9369 (t = -5.78). A 1% rise in the index, 0.01 in log terms, times -0.9369 is a -0.94% log-gold move. Keep that one. Drop the real-yield duration from about 7.3 years toward about 1.3, and re-estimate it when the post-2022 window is longer than 206 weeks. Inside the window a Chow test, a break test at a date you name, is already marginal around December 2023 (F = 2.62, p = 0.049). The third regime fails a break test on its own dates.

I am against two uses of the residual. Keeping the full-sample monthly slope of -0.064 on a book that has lived through 2022 averages a window where the weekly slope was -0.073 with a window where it was -0.013\. And filling the gap with official-sector demand, frozen reserves, or a fiscal term premium treats three hypotheses the authors list as the estimate. They say so: the bivariate regression cannot assign the residual, weekly central-bank purchases are not in the data, and IMF reserve holdings arrive quarterly. Until that flow is a regressor, the coefficient you can defend is about zero on the real yield and about -0.94 on the dollar.

## KEY POINTS

- Estimate the log change in the London gold fix on the change in the 10-year TIPS yield, in percentage points, and the log change in the ICE dollar index. The monthly TIPS slope is -0.0639: 100 basis points, dollar unchanged, is a 6.39% drop in log gold, a 6.19% drop in the price.
- Split the months at a TIPS yield of zero. Below zero the slope is -0.0912 on 49 months; above zero it is -0.0625\. The weekly Hansen kink is 1.78%, and the monthly threshold test does not reject (p = 0.114). The slopes at 1.78% are not printed.
- Inside the zero-rate years, re-estimate on non-overlapping horizons. The TIPS slope goes from -0.073 at one week to -0.188 at 26 weeks, and that 26-week row has 26 observations.
- After 25 March 2022, use the weekly TIPS slope of -0.0125 (t = -0.98). The full-sample monthly slope of -0.064 averages that window with the zero-rate years. The intercept is +0.50% a week, which they annualize to +26% by multiplying by 52.
- Do not trade the lag. The lagged monthly mean-squared-error ratio is 1.03 times the historical mean. Hedge the dollar, which still loads at -0.94, and leave official-sector buying out of the hedge ratio until it is in the regression.

## References

- [Why FX Traders Must Watch Gold, Rates, and Equities](https://aligrithm.com/why-fx-traders-must-watch-gold-rates-and-equities/)
- [Stock-Bond Correlation: The Sign Flips and So Do Its Drivers](https://aligrithm.com/stock-bond-correlation-the-sign-flips-and-so-do-its-drivers/)
- [Stocks and Bonds Predict FX — But Only in Emerging Markets](https://aligrithm.com/stocks-and-bonds-predict-fx-but-only-in-emerging-markets/)
- [Gold and U.S. Treasuries as Competing Safe Assets, Jonathan A. Batten, Igor Loncarski, Peter G. Szilagyi, and Hanxian Zhou](https://papers.ssrn.com/sol3/papers.cfm?abstract%5Fid=6627301&ref=aligrithm.com)

---

**A note on AI.* The ideas, research, analysis, and conclusions in this article are my own. I use AI tools to help with editing and wordsmithing, because English is not my first language, and I am not shy about that. AI-generated ideas and AI-assisted writing are not the same thing: the first is empty slop from a generic prompt, the second is a tool for communicating years of real research more clearly. Judge the work by its substance, not by whether software helped polish the prose.*