There is something paradoxical about Michael Gove's recent speech calling for government to be "rigorous and fearless in its evaluation of policy and projects." It's that his praise for evidence-based policy has come in a year when we've seen that policy should sometimes not be based on rigorous evidence.

The best time to have imposed the lockdown was as soon as possible after a few cases had been discovered. But this would have been a hard sell. Not just the grifter media but the public would have asked: why are we losing our freedoms to forestall so small a threat? But of course, by the time there was strong evidence that a lockdown was necessary, it was too late.

Tens of thousands of us face the problem of having to act on insufficient evidence. The equity investor who waits for strong evidence that a share is under-priced will buy when it no longer is. Loads of hedge funds exemplify Gove's ideal: they employ lots of data scientists working with strong incentives and clear metrics. But their results are generally mediocre. Similarly, the entrepreneur who waits for proof that there's a market for a product will find that his rivals have captured it. The factory manager who waits for good evidence that a machine needs replacing will see it blow up. And so on.

The same applies to macroeconomic policy. The Bank of England did not cut Bank rate to 05% and introduce Qe until March 2009 - 12 months after the recession had begun. Waiting for strong evidence of a recession before loosening policy means loosening too late. This is why I favour strong automatic stabilizers.

In these cases, the demand for rigorous evidence is an example of what Jon Elster has called hyper-rationality:

Sometimes people ignore the costs of decision-making. They search for the solution that would have been best if found instantaneously and costlessly, ignoring the fact that the search itself has costs that may detract from optimality. (Solomonic Judgements, p260)

There are other problems. It's sometimes hard to extrapolate the results of the RCTs lauded by Gove. They "prove", for example, that parachutes don't save lives. As Cartwright and Deaton say:

Demonstrating that a treatment works in one situation is exceedingly weak evidence that it will work in the same way elsewhere; this is the 'transportation' problem.

And where they do yield results, these can be hard to interpret. The average treatment effect hides important heterogeneity of effects on individuals, for example.

There's also the problem that past evidence is no guide to the future. We have evidence from the 50s and 60s that very high top tax rates are entirely consistent with strong GDP growth. But is this evidence useful? Mr Gove and his friends would probably say not. And perhaps for a good reason: high taxes are sustainable if those paying them regard them as a price worth paying to live in a civilized society, but resist them if they regard them as infringements of their rights. Policies effects depend upon beliefs. And these change.

Indeed, they sometimes change because of policy. One effect of the expansion of universities in the 90s has been to create a large cohort of metropolitan liberals. How do we factor this into cost-benefit analyses?

There are two very important categories of cases where past evidence doesn't help us predict the future.

One is the problem of radical uncertainty: we can never be 100% confident that past statistical relationships will continue to hold. The other is that of reflexivity. Beliefs can change the world. For example, McLean and Pontiff and Cotter & McGeever have shown that when strong evidence emerges of stock market anomalies, these get competed away. In economic policy, the counterpart to this is Goodhart's law: "any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes." You can't solve these problems with big data: the datasets used to find stock market anomalies are huge, but their findings less so .

All of which leads us to three paradoxes.

Paradox one is that, as Abby Innes points out, there is a precedent for Gove's belief that data science will deliver better government. The old Soviet Union had a "fantasy of 'optimal government'" in which "cyberneticians would depict the governmental system as an object of technical control, with inputs, outputs, and feedback loops: the language of machines." This is evoked in Francis Spufford's wonderful Red Plenty. In this sense, Gove's role model is not so much FDR as Leonid Kantorovich. This project, however, failed in part for the reason Hayek pointed out - that governments just cannot know enough about complex systems.

Paradox two, as Phil points out, is that Gove's vision is unconservative. It is one in which:

the centrality of government is overweening and checks on the executive's remit - other state institutions, watchdog quangos, parliamentary scrutiny and media accountability - are brushed aside by authoritarian, action-oriented government.

Which leads to paradox three. One the question: is data-driven effective government likely, it is the conservative Gove here who is taking the side of the old USSR and this Marxist who is one the side of Hayek.