Oil supply, demand and risk shocks

Download the data and code

• oil_shocks_monthly.csv488 months · 62 KBOne row per month: WTI price, oil price change, VIX innovation, the three shocks, their cumulative sums, realtime versions of the shocks and an NBER recession flag.
  Download CSV
• update_oilshocks.py10 KBPython script that downloads the inputs and rebuilds the file. Needs numpy, pandas and scipy.
  Download Python
• update_oilshocks.m10 KBMATLAB script that does the same. No toolboxes needed.
  Download MATLAB
• replicate_ready.py3 KBRebuilds the shocks over the paper’s 1986 to 2011 sample and prints them beside the paper’s Tables III and IV.
  Download

Both scripts download the inputs when they run and estimate the ARMA(1,1) by the same conditional least squares, so they give the same output to the six significant figures written in the file. The whole series is re-estimated each time: see the notes below.

Read the code

Loading the script…

How the shocks are identified

Three monthly series go in: the log change in the oil price, Δp; the log return on an index of oil producers, R^prod; and the innovation to the VIX, ξ_VIX, the residual of an ARMA(1,1) fitted to the log of the index. Ready (2018, Section 3.1) writes the observed vector X = (Δp, R^prod, ξ_VIX)′ as X = AZ, where Z = (s, d, v)′ holds the supply, demand and risk shocks and

A = 1110a₂₂a₂₃00a₃₃

The first row says the three shocks add up to the oil price change. The zeros say that producer returns do not respond to supply shocks, and that the VIX innovation responds only to the risk shock. The remaining three elements of A are chosen so that A⁻¹Σ_X(A⁻¹)′ is diagonal, where Σ_X is the covariance matrix of X: the shocks are uncorrelated by construction. This is the usual recursive identification of a structural VAR, renormalised so that the shocks are measured in units of the oil price rather than with unit variance. In regression terms, v is the fitted value of Δp on ξ_VIX; d is the fitted value of Δp on the part of R^prod that is orthogonal to ξ_VIX; and s is the residual.

Two details are choices made here. The demand and risk shocks are given a mean of zero, so the average drift in the oil price, 0.3% a month since 1986, sits in the supply shock; the covariances, and so A, are the same either way. And the ARMA and Σ_X are estimated over the whole sample, as in the paper, so every value in the file is re-estimated when a month is added. The _realtime columns hold the shocks estimated only on data up to each month, from a first window of 120 months; they never change once published, and their correlations with the full-sample shocks are 0.99 for supply, 0.99 for demand and 0.72 for the much smaller risk shock.

Replication

Before extending anything, the decomposition was rebuilt over the paper’s own sample, January 1986 to December 2011, from the public inputs below. The paper’s summary statistics (Table III, Panel B) and its regression of the US market return on the shocks (Table IV, Panel A) are shown beside the same figures computed here. replicate_ready.py repeats the comparison.

Market return: the CRSP value-weighted index in the paper, Fama and French’s market return (Mkt-RF plus RF) here; both in logs with White standard errors. The paper’s regressions use 315 months, the rebuild 311 (February 1986 to December 2011).

The paper’s central result survives the change of inputs: on its own the oil price change has no relation to the stock market (R² of 0.3%), but once it is split, demand shocks are strongly positive for returns and supply shocks strongly negative. The industry pattern in the paper’s Table V also holds with the Fama–French ten industries: consumer durables, retail, non-durables and the “other” group load most negatively on supply shocks, energy not at all, and every industry loads positively on demand shocks.

Inputs and what changed

What the shocks say about past episodes

Sums of the monthly log changes over each episode, in log points. The Gulf War is the paper’s own test case (its Section 3.3.a): a supply shock that oil producers’ shares shrugged off while the market fell.

Before you use it

• Everything is re-estimated each month. The ARMA(1,1) and the covariance matrix use the whole sample, as in the paper, so adding a month changes every earlier value a little. Use the _realtime columns if you need shocks that were knowable at the time, and note that they start in December 1995, the end of the first 120-month window.
• Timing. A shock for month t uses the oil price, producer return and VIX at the end of t. The producer return comes from Kenneth French’s library, which is updated with a lag of a month or two, so the latest month is published later than the oil price itself.
• Spot, not futures. The oil price change is the month-end spot WTI. It is more volatile than the futures return the paper used, and in April 2020 it reflects the Cushing storage squeeze (the month-end price stayed positive; the negative print was on 20 April).
• US producers, not world. The producer index covers US-listed oil firms, which include the big integrated majors but not state producers, as the paper’s index also did not.
• Means. The demand and risk shocks average zero; the supply shock carries the oil price’s average drift. If you want all three demeaned, subtract the mean of s.
• Units. Log changes as decimals in the CSV (0.01 = 1%); log points × 100 in the chart and the tables on this page.
• Data terms. The file holds the oil price (public, from the EIA via FRED), the VIX innovation and the shocks. The producer returns themselves are not redistributed; the scripts download them from Kenneth French’s library when they run.

Why it matters

Oil price changes and stock returns are almost uncorrelated, which has long looked like a puzzle given how much oil matters for the economy. Ready (2018) resolves it: supply-driven rises in the oil price are bad for stocks and for output, demand-driven rises are good for both, and the two cancel in the raw data. Because the shocks are identified from prices rather than from production data or news events, they are available monthly and in real time, which makes them usable in forecasting and asset-pricing work that quarterly SVAR shocks cannot support.

How to cite

If you use the shocks, please cite the paper that defined the method. A reference for this page is included in case it helps.

@article{Ready2018,
  author  = {Ready, Robert C.},
  title   = {Oil prices and the stock market},
  journal = {Review of Finance},
  volume  = {22},
  number  = {1},
  pages   = {155--176},
  year    = {2018},
  doi     = {10.1093/rof/rfw071}
}

@misc{PybisOilShocks,
  author       = {Pybis, Sam},
  title        = {Oil supply, demand and risk shocks after {Ready} (2018), monthly, 1986 to August 2026},
  year         = {2026},
  howpublished = {\url{https://sampybis.co.uk/updated-data/oil-shocks.html}},
  note         = {Updated 25 October 2026}
}