The Price of Knowing

Two numbers, two kinds of knowledge

Begin with the diesel price itself.

EIA's weekly retail price is a statistical measurement. Its current methodology uses a stratified sample drawn from a frame of roughly 73,000 service stations and 9,500 truck stops. Prices are collected as of 8 a.m. each Monday. EIA says responding outlets represent at least 80% of weighted national diesel sales; missing observations may be imputed from an outlet's previous response, comparable outlets and commercial information. It publishes its methodology and measures of sampling error.¹

It is difficult for an administration simply to wish such a number away. Private information provides an external check. GasBuddy's measure is not directly comparable because it is not the same sample, collected at the same moment and weighted in the same way. Even so, its crossing $6 on September 10 is consistent with the direction and scale of EIA's Monday observation.²

The forecast is different.

EIA's September outlook expects diesel to average $5.55 in the fourth quarter of this year and then decline to $4.40 during 2027. That projection depends on an elaborate view of the future petroleum system. EIA expects Middle Eastern oil production to increase as more petroleum moves through the Strait of Hormuz and alternative export routes expand. It assumes regional crude production will move towards pre-conflict levels by the second quarter of 2027. With production recovering and inventories eventually rebuilding, it expects Brent crude to average about $74 next year.³

These are not absurd assumptions. They may turn out to be correct.

But they are assumptions.

EIA is fairly explicit about their importance. Its petroleum-products analysis says the projected decline in diesel refining margins assumes a return towards normal tanker traffic and greater availability of Middle Eastern crude. If constraints persist beyond the end of 2026, EIA says global distillate crack spreads would be higher than forecast.³

That conditional sentence arguably contains more useful information than the headline number of $4.40.

The world confronting the forecast is not especially reassuring. EIA expects American distillate inventories, including diesel and related fuels, to fall below 100 million barrels and remain below their recent five-year range through much of 2027. Private-market data cited by Reuters on September 10 put American diesel inventories about 13% below their five-year average and the diesel crack spread at a record $112.17 a barrel.²

None of this proves EIA wrong. It demonstrates something more mundane and more important: a forecast is perishable.

The six-day-old future

The September outlook was published on September 9. EIA states prominently that it finalised the inputs to its model on September 3 and that the forecast does not specifically account for events after that date.³

Every serious forecasting exercise requires a cutoff. Data must eventually stop entering the machine so that models can run, analysts can review them, tables can be produced and publications can be prepared. The existence of an input date is not suspicious.

The interesting question is what should happen when reality changes materially between the cutoff and publication—or immediately afterwards.

That is a governance question, not an accusation. What constitutes sufficient materiality? Who makes the decision? Is there a public rule, or only professional judgment exercised behind closed doors?

The energy system is being buffeted by an active Middle Eastern war, disrupted Gulf transportation, attacks on Russian refining capacity, low American distillate inventories and historically high diesel refining margins. EIA itself substantially revised the outlook between August and September: its forecast for 2027 diesel rose from $4.07 to $4.40, an 8.1% increase in a single month.³

There is therefore no evidence that the agency simply ignored adverse information. Quite the opposite. It raised its forecast, disclosed the model cutoff and described the risks to its baseline. When it discovered that the initial release had mistakenly displayed gasoline crack-spread figures as distillate figures in one table, it issued a correction the following day.⁴

Those facts matter. They are evidence against the crudest theory of political manipulation.

They do not resolve the subtler question of whether the institution's baseline, update rules and presentation of uncertainty are adequate to the environment in which it operates.

The politics of the baseline

Models cannot forecast without a central case. Unfortunately, a “baseline” sounds far more objective than it really is.

Consider a model attempting to forecast diesel prices during a war. Analysts must make judgments about when damaged infrastructure returns, when tankers resume normal routes, how quickly producers restore output, whether attacks expand or diminish, how governments respond, whether inventories are released, whether demand weakens and how refiners change their product mix.

None of those judgments requires partisan intent. All of them influence the answer.

A baseline in which Gulf production normalises relatively quickly will generate one price path. A baseline in which disruption persists for another year will generate another. Both can be internally coherent. Both can be generated by respectable economists. The important question is why one becomes the forecast while the other becomes merely a risk.

This is where political economy enters the forecasting room even when politicians do not.

The executive branch has an obvious interest in lower energy prices. Every administration does. Fuel prices are visible, frequently paid and politically salient. In 2026 they are particularly awkward because the supply shock is connected in part to a conflict in which the United States is directly involved. The White House has separately invoked the Defense Production Act to expand domestic petroleum production, refining and logistics capacity, describing current constraints as a threat to prosperity and national security.⁵

None of that means EIA analysts altered their model to help the White House. It means something more basic: the political value of a reassuring baseline is unusually high.

An institution designed to produce independent knowledge should be strongest precisely when the government surrounding it has the greatest reason to prefer a particular answer.

Independence can be starved

Congress anticipated the problem of political influence when it created EIA. The agency is not structurally independent in the manner of the Federal Reserve: it sits inside the Department of Energy, and its Administrator is appointed by the President with the advice and consent of the Senate. But it enjoys unusually explicit statutory protection.

Under 42 U.S.C. § 7135(d), the Administrator is not required to obtain another DOE official's approval when collecting or analysing information, and does not have to secure another federal officer's approval before publishing the substance of statistical or forecasting technical reports.⁶

This architecture reflects a broader principle of official statistics. The National Academies' 2025 Principles and Practices for a Federal Statistical Agency says federal statistical bodies must be independent of political and other undue external influence when developing, producing and disseminating statistics. It identifies control over content, methods, frequency and release, without prior political clearance, as necessary protections.⁷

The reason is not aesthetic. The product being manufactured is credibility.

An oil trader can observe futures prices. A household can see what it pays at the pump. But hardly any individual citizen, business or investor can reproduce the government's statistical machinery. Official statistics therefore depend on trust. Once users apply a partisan discount to every official number, even accurate statistics become less valuable.

Legal independence is necessary. It is not sufficient. A statistical agency also needs statisticians.

DOE's own budget documents show 353 federal EIA positions in fiscal 2025 and 246 in fiscal 2026. That is a decline of 107 posts, or about 30%. The department's fiscal 2027 request keeps the agency at 246 full-time-equivalent positions.⁸

Staff reductions do not prove political capture. Governments set budgets, and an agency can sometimes become more efficient. But there is no serious conception of statistical independence in which human capacity is irrelevant.

The National Academies treats the recruitment and retention of qualified professional staff, methodological expertise, quality assurance and continued technical improvement as parts of statistical integrity.⁷ A legislature can grant an agency formal authority to decide its methodology and still leave that authority increasingly theoretical if the agency lacks enough experienced people to maintain surveys, improve models, test alternative assumptions and respond quickly when the world changes.

This is one path by which institutional integrity deteriorates without a censor ever arriving from the White House.

Capture without conspiracy

Economists normally associate “capture” with regulation: the political process through which concentrated interests bend public power towards themselves. A statistical institution presents a different problem. It does not principally tell companies what they may do. It tells society what is happening.

Its valuable asset is epistemic authority.

The relevant danger might therefore be called epistemic capture: a process in which the machinery that determines how reality is measured, modelled and communicated becomes aligned with a narrower set of incentives or assumptions. Recent scholarship uses the term to describe how institutions can privilege particular frameworks of expertise until those frameworks constrain what is considered knowable or reasonable.⁹

For a government forecasting agency, capture need not look like a political operative ordering analysts to lower next year's diesel price. It can take less theatrical forms: an alternative scenario is no longer produced because nobody has time to maintain it; a normalisation assumption persists because changing it requires a lengthy review; a release schedule becomes more important than the freshness of its inputs; politically awkward forecasts attract more internal scrutiny than convenient ones; technical caveats survive in the report but vanish from its public meaning.

Each decision can be individually defensible. Together they can change the epistemic character of an institution.

That is what makes the idea more unsettling than ordinary falsification. Fraud creates a false number. Epistemic corruption can create a perfectly reproducible number through a process that has gradually become less capable of discovering when its assumptions are wrong.

This concern does not arise in a vacuum. On August 1, 2025, President Trump fired Erika McEntarfer, the Commissioner of the Bureau of Labor Statistics, immediately after a disappointing employment report and large downward revisions to earlier estimates. He accused her of manipulating the figures to damage him politically, but presented no evidence. Economists and former officials warned that the dismissal could damage confidence in American statistics.¹⁰

The episode matters even if no EIA employee ever received a political instruction. Institutions learn from what happens to neighbouring institutions. If senior statistical officials can lose their jobs immediately after releasing unwelcome information, career analysts elsewhere do not need an explicit threat to understand that political consequences exist.

This does not establish that EIA changed its work. It does reinforce why statistical independence must ensure that analysts never have to ask whether an inconvenient conclusion will have consequences.

The peculiar economics of diesel

Why devote so much attention to a fuel forecast? Because diesel is unusually close to being an economy-wide input.

America's trucks moved 65% of domestic freight by weight and 72% by value in 2023, according to the Federal Highway Administration.¹¹ The American Transportation Research Institute calculates that the average cost of operating a truck reached a record $2.336 per mile in 2025. Excluding fuel, it was $1.854, implying roughly 48 cents per mile of fuel cost at last year's prices.¹²

The 2025 national diesel average was $3.660. The latest EIA reading of $5.967 is about 63% higher. It is nearly 20% above even 2022's full-year average of $4.989.¹³

A crude illustrative calculation is sobering. If a carrier's fuel cost per mile simply increased in proportion to the national pump price, 48 cents would become roughly 79 cents. Holding everything else constant, total operating cost would move from about $2.34 to roughly $2.64 a mile. That is an increase of around 13%.

Real fleets do not operate so mechanically. They negotiate discounts, impose fuel surcharges, change routes and improve utilisation. But somebody ultimately bears most of the cost: a shipper, a farmer, a builder, a retailer or a consumer.

Modern research generally finds that oil shocks have a more modest effect on underlying inflation than the experience of the 1970s might suggest. A Federal Reserve model of the 2022 oil shock estimated that the rise in prices added almost one percentage point to headline inflation on impact, while the effects on core inflation and output were much smaller.¹⁴

History still counsels against treating energy shocks as trivial. James Hamilton's work on the 2007–08 oil-price surge concluded that it materially reduced consumption and vehicle purchases and contributed to the initial recessionary dynamics, even though the financial crisis ultimately dominated the downturn.¹⁵

The current problem has an additional wrinkle: diesel itself is unusually tight. This is not merely expensive crude passing cleanly through a refinery. Distillate inventories are low, global refinery production has been disrupted and diesel crack spreads have become extreme.²³

Whether those conditions disappear in six months or eighteen is economically consequential. It is also politically consequential. That is precisely why the institution making the forecast deserves unusual scrutiny.

A model is not a promise

There is a common failure in public discussions of forecasts. An agency publishes a conditional central estimate. A press release reduces it to a sentence. Politicians and journalists convert the sentence into a factual proposition.

A model assumption becomes an expectation. The expectation becomes a promise.

Thus:

EIA forecasts diesel at $4.40 in 2027.

quietly becomes:

Diesel will fall to $4.40 next year.

Those sentences are not equivalent.

The first means something approximately like this: given EIA's current assumptions about production, transportation, refining, demand and geopolitical conditions, the model's central annual estimate is $4.40.

The second sounds like knowledge.

One way to improve institutional integrity would be to make that distinction harder to erase. EIA already discloses its assumptions, and the September report deserves credit for doing so. But extraordinary geopolitical periods warrant more than ordinary disclosure.

A more robust forecasting regime could publish at least two explicit alternatives alongside its baseline: a faster-normalisation scenario and a persistent-disruption scenario. Rather than leaving geopolitical uncertainty in prose, it could show how $4.40 changes if Gulf flows do not substantially normalise by the second quarter of 2027.

It could attach a conspicuous “information as of” date to every headline forecast and establish public materiality criteria for rerunning a forecast after the normal input cutoff.

It could publish retrospective forecast errors according to their source: demand, production, geopolitical assumptions, refinery outages and model misspecification. During extraordinary shocks, it could issue interim scenario updates without pretending to generate a completely new central forecast.

These measures would not make EIA omniscient. They would make its uncertainty legible.

Has EIA been captured?

On the evidence presently available, that conclusion would go too far.

Important facts point in the opposite direction. EIA has strong statutory protections. It continues to publish politically inconvenient information: record fuel prices, depleted inventories and severe Middle Eastern supply constraints. It raised rather than reduced its diesel forecast this month. It disclosed the September 3 input freeze, warned that longer-lasting Gulf disruption would raise diesel margins beyond its baseline, and corrected an error one day after publication.³⁴⁶

These are not the obvious behaviours of an agency engaged in crude propaganda.

Yet there are legitimate reasons for heightened scrutiny. EIA exists inside an executive department whose political leadership is appointed by the President. Its Administrator is a presidential appointee. The federal statistical system has already experienced an extraordinary political intervention in the firing of the BLS Commissioner after an unfavourable report. EIA has undergone a large reduction in federal staffing. And its most politically salient current forecast depends on a geopolitical normalisation assumption concerning a conflict in which the government that houses the agency is itself a participant.

None of those facts proves corruption. Together they mean the integrity of the process matters more than usual.

The right response is neither credulity nor conspiracy. It is verification.

Statistical agencies enjoy a credibility premium. Markets can act on an EIA inventory report because traders generally assume the government has not constructed it to help the President. Businesses can write contracts around BLS inflation statistics because both sides accept the measurement infrastructure. The Federal Reserve can cite government employment data because investors believe the numbers were generated before anyone knew whether they would flatter the incumbent party.

Destroy that assumption and the price of information rises everywhere. Every release requires political interpretation. Every revision becomes suspicious. Private alternatives proliferate, but not evenly: large companies and hedge funds can buy proprietary freight, tanker, satellite and credit-card data. Ordinary businesses and citizens remain dependent on public statistics.

Weakening official statistics does not eliminate information. It makes good information expensive.

That is why institutional integrity cannot be reduced to the question: Did anyone falsify the number? That is the lowest imaginable standard.

A statistical agency worthy of a sophisticated economy must answer harder questions. Did we use the best information reasonably available? Did we update when circumstances materially changed? Did we stress-test the assumptions most likely to fail? Did professional staff have adequate resources? Could analysts reach an inconvenient conclusion without worrying about political consequences? Did our public communication convey the conditionality of the model rather than merely its most reassuring number?

And would we have made the same methodological choices if the resulting forecast were politically inconvenient to a different administration?

Those are questions of process. They are also questions of power.

Reality eventually marks the model to market

Energy forecasters have one advantage over many political institutions: petroleum markets are unforgiving.

If diesel remains near $6 next year, no press release will make it $4.40. If tanker traffic remains constrained, inventories will reveal it. If refineries cannot obtain crude, crack spreads will reveal it. If freight costs rise, trucking accounts will reveal it.

Reality will eventually mark the forecast to market.

The danger lies in what happens before then. Governments make decisions using forecasts. The Federal Reserve forms expectations. Businesses budget. Voters evaluate policy. An optimistic baseline can shape behaviour during the period when its accuracy remains unknowable.

This is why the epistemic integrity of forecasting institutions is not an academic concern.

In an authoritarian state, the most obvious statistical pathology is fabrication: the ruler dislikes the number, so the number changes. A modern bureaucracy offers more sophisticated possibilities. Nobody needs to falsify anything. Models can remain technically competent. Analysts can remain conscientious. Every procedure can be defensible in isolation.

Yet a government can still drift towards a world in which the official production of knowledge becomes systematically less capable of contradicting the assumptions on which the government would prefer to operate.

That is institutional capture in its quieter form. And that is what makes epistemic corruption dangerous.

The spreadsheet may remain immaculate.

The danger begins when an institution forgets that its first obligation is not to the model, the release calendar, the department or the administration. It is to reality.