A positioning variable that leads gold prices by ten months would be among the most valuable single facts in commodity trading. Başaran's abstract puts the lead at "approximately ten months" and describes it as "statistically significant and deterministic". Yet the evidence comes from the peak of a recurrence statistic, calculated once on one sample of 233 monthly observations, using one radius setting and without an attached significance test.
Gold pays no coupon, and real rates determine its carry. Higher real rates raise the opportunity cost of holding a non-interest-bearing asset, creating the textbook expectation of downward pressure on price. Başaran asks which side moves first. Real interest rates represent the public sector, while the long-short spread on one-month gold options represents the capital market. The paper therefore frames gold prices as the object of a contest between public and private sector influence.
The main method is Cross-Recurrence Quantification Analysis (CRQA), which counts how often two systems move through similar states. Başaran applies it to two pairs, then repeats the exercise in rolling form with a 48-month window stepped every 2 months. The monthly sample runs from January 2006 to November 2025, N = 233. It contains real gold price (gp), the US 10-year borrowing rate deflated to a real rate (ri), the one-month gold option long-short difference (gs), and the Dollar Index (di).
Two hypotheses appear up front. H1: speculative positions are a long-term leading indicator of gold prices. H2: gold prices are an early warning indicator that drives real interest rates with long lags.
Both are reported as confirmed. The rates and gold pair has determinism of 74.56% and a lag of -21 months, interpreted as gold leading rates. The speculation and gold pair has determinism of 56.68% and a lag of +10 months, interpreted as speculation leading gold.
What the instrument measures
CRQA embeds each time series in a phase space. It marks each pair of times (i, j) for which the trajectories fall within distance epsilon of one another, producing a cross-recurrence plot. The method then counts structures in that grid. Recurrence rate (RR) is the share of the grid containing marks. Determinism (DET) is the share of those marks found on diagonal lines, while laminarity (LAM) is the share contained in vertical structures. ENTR measures the entropy of diagonal line lengths.
Diagonal lines indicate that the systems pass through similar sequences of states at some offset. Summing the marks along each diagonal produces the cross-recurrence profile, DRP(tau). Its highest point is Pmax, and the sign of that lag determines the stated direction of the lead.
Başaran runs this procedure on the monthly sample above. Real gold price is nominal gold deflated by a dollar inflation index. The speculative series is described as the difference between short and long position values of one-month gold options. In the data section, the Dollar Index has mean 106.146 and SD 12.467. I could not find it entering either CRQA test.
The treatment of di raises another problem. The data list calls it the Dollar Index (DXY), whereas the setup says real gold prices were calculated with the dollar inflation index (di). DXY is a currency basket. The deflator used to construct gp therefore differs from the object named in the variable list.
For ri and gp, the reported settings and results are radius 0.13, RR 3.21%, DET 74.56%, ENTR 1.43, LAM 80.33%, and Pmax at -21 months. The paper reads this as gold leading rates. For gs and gp, they are radius 0.15, RR 1.99%, DET 56.68%, ENTR 0.91, LAM 77.30%, and Pmax at +10 months, read as speculation leading gold.
Using the 48-month window and 2-month step, the analysis produces average DET of 39% for gp-gs and 58% for gp-ri. Both fall to zero in 2008. gp-gs also reaches zero in 2020, while gp-ri does so in 2022.
Başaran's economic interpretation is sweeping: capital markets price gold ten months in advance, the state responds twenty-one months later, and most chaos in gold comes from internal capital dynamics, with policy serving as a secondary source.
Which series represents positioning?
The paper's data list defines gs as speculative positions in gold, measured through a long-short difference. In the results tables, the same series becomes "gold option spreads". The setup describes it as "the difference between the short and long position values of one-month gold options". These labels refer to different objects. A long-short open-interest difference measures positioning. A spread between option values is a price.
gs has mean 1068.676 and SD 661.253, figures that do not resolve the ambiguity. I could not find a stated data source for any of the four series, including gs.
The paper acknowledges the dispute over how much information speculative positions contain. Its conclusion quotes both sides directly. Wang (2002) suggests that speculative positions may be associated with market instability. Aggarwal and Thomas (2018) show that speculative trading can play a central role in information production under certain market conditions.
The meaning of a +10 month lead depends on what gs measures. If gs is an option price spread, the paper has found that one gold-linked price leads another gold price by ten months. Anyone trying to check that result would need the contract definition. Neither the contract nor the vendor is named.
We could not run this ourselves. The central input is a speculative positioning feed, and we hold no gold derivatives-positioning data. An exercise using rates and gold alone would remove the paper's main variable and answer a different question.
A DRP peak is still some distance from a forecast
The +10 peak records pattern overlap between two embedded trajectories. The paper never tests whether today's gs narrows the distribution of gp ten months ahead.
The series enter in levels and are z-scored, without differencing or detrending. Başaran gives a direct defence, citing Marwan et al. and Webber and Zbilut: traditional linear smoothing operations such as differencing behave as high-pass filters and may distort attractor geometry. The recurrence literature has a real argument here.
The consequence is equally real. Gold contributes a large common trend to recurrence structures shared with another trending series. Its sample mean is 1494.173, with SD 587.952 and skewness 1.487. Two series drifting in the same direction can generate long diagonal lines. DET of 74.56% between rates and gold could reflect genuine coupling. It is equally compatible with two slow-moving levels series.
A surrogate test could distinguish those cases. Shuffled or phase-randomised series would preserve the trend while destroying the dynamics. I did not find such a test in the paper.
The abstract describes the speculative lead as "statistically significant and deterministic." I found no test statistic, no p-value and no bootstrap distribution supporting the significance claim. DET is a percentage.
One target band determines the radii
Başaran states the radius procedure plainly and quotes its logic: an epsilon set too wide makes everything look similar through saturation, while one set too narrow misses real links through sparsity. The target is RR between 2% and 5%, attributed to Webber and Zbilut and Marwan et al. A radius of 0.13 yields RR 3.21% for one pair. For the other, radius 0.15 yields RR 1.99%. The results sit inside or on the edge of the intended band because the radii were chosen for that purpose.
Those radii are fitted in-sample to a target using the same 233 observations that generate the headline results. As a convention, that is fine. The concern is that epsilon changes both DET and the shape of DRP. I did not find a sensitivity table showing where Pmax lands at 0.10 or 0.20.
The embedding receives no corresponding sensitivity table. Table 2 gives Embedding 3, Dimension 1, Lag 1 for both pairs, an ambiguous presentation as printed. The paper says parameter selection, covering embedding dimension, lag and radius, was optimized separately for each variable pair to keep RR within the 2-5% band. It supplies no procedure for choosing the embedding and no sensitivity check for either pair. Since embedding choices determine what qualifies as a recurring state, the stability of the peak matters. The entire result rests on its location.
A smaller inconsistency appears nearby. The method section defines ENTR as the complexity of the dynamic structure. In the discussion of Table 4, it becomes the average of the diagonal line lengths. That quantity differs from entropy.
The rolling evidence moves the lag
Başaran makes the concession explicitly. The conclusion says that "leadership is not fixed, but weakens or strengthens depending on time and market conditions" and that "the information leadership of speculative positions is neither continuous nor unconditional". The abstract agrees: "time-dependent analyses reveal that these relationships weaken during periods of crisis and uncertainty and temporarily dissolve in some sub-periods".
That leaves an obvious tension with the same abstract's decision to report approximately ten months as a statistically significant figure. The conclusion argues that speculation "can exhibit a highly systematic behavior under appropriate market regimes". No ex ante rule identifies those appropriate regimes anywhere in the paper. Without one, the regime concession cannot preserve the ten-month figure as a trading signal.
Table 6 divides the sample into epochs. During 2009-2012, gs leads gp at +20 months. In 2011-2013, ri leads gp at +18, at peak connection strength. The direction changes over 2013-2016, when gp leads gs at -18. Across 2014-2021, gp leads ri at -21. Synchronization reaches its peak in April 2019. The paper calls 2022-2025 indecisive, with very low Pmax.
Both pairs reverse direction. Within the speculation pair, the lag spans +20 to -18 months. The full-sample +10 combines epochs that conflict with that figure and with one another.
Determinism shifts as well. Average windowed DET for gp-gs is 39%, below the full-sample 56.68%, and reaches zero in 2008 and 2020. For gp-ri, the windowed average is 58%, compared with 74.56% for the full sample, with zeros in 2008 and 2022.
The 2022 zero deserves attention. An opportunity-cost relation between real rates and gold should be most visible during a rapid hiking cycle. Here the recurrence structure disappears.
Each window covers 48 months and advances by 2 months, leaving consecutive readings with 46 months of shared data. These regime "shifts" are heavily overlapping views of the same observations. A sequence of shifts can arise when one or two months enter the window.
The conclusion also contains a contradiction. During crises, the paper says, "the public manages gold prices and the state becomes an actor that controls prices." Its discussion identifies the 2008 crisis and the 2020 pandemic as periods when the system's deterministic structure temporarily disappeared. Determinism reaches zero in 2008, then does so again for the rates pair in 2022. Zero determinism means the structural link is absent in either direction.
The part that survives
The reversal involving rates is the paper's genuinely useful finding, and Başaran states it as a hypothesis in advance. H2 says gold prices drive real interest rates with long lags. The abstract highlights the result by describing public policies as responding "in a lag and largely reactive manner". Real rates adapting to gold with a delay reverses the textbook direction. A claim of that size needs a method that supplies standard errors.
The regime evidence also carries weight insofar as it shows that any coupling is episodic. Average windowed determinism of 39%, including zeros in two crises, offers little basis for sizing a book against the relationship.
The ten-month lead does not survive as a signal. The paper presents no strategy, no return series, no Sharpe and no cost assumption. Its own 2013-2016 window moves the DRP peak to -18. With a monthly series containing 233 points, a ten-month forecast horizon leaves roughly 22 non-overlapping observations of the claimed prediction target. Trading costs are absent because the paper provides nothing against which to charge them.
One result would change my mind: surrogate-data nulls showing DET and Pmax outside the distribution generated by trend-preserving shuffles, alongside a radius sweep in which the same Pmax remains in place. Until then, the paper describes pattern overlap between two trending levels series and assigns a lag to it.