11Intellectual Property and Non-Competes

On 5 June 2009, his last day at Goldman Sachs, a programmer who had worked on the bank’s high-frequency trading system encrypted and uploaded more than 500 000 lines of its source code to a server in Germany; he had accepted a job at a start-up building its own trading system. A jury convicted him under two federal statutes and he was sentenced to 97 months in prison. In 2012 the Court of Appeals for the Second Circuit reversed both convictions: the code, it held, was not a stolen “good” under one statute, and under the other it was not related to a product “produced for or placed in interstate or foreign commerce”, since the bank sold and licensed its trading system to no one. Within the year Congress amended the second statute to cover products and services used in commerce. A trading firm’s most valuable assets are written in code and carried in people’s heads, and the law that protects them turns on words like these.

11.1 What a trading firm owns: code, models, data and know-how

A trading firm’s capital, premises and connections can be bought. What it owns that a competitor cannot buy is a set of things that are hard to see and easy to carry: source code, the structure of its models and their fitted parameters, cleaned and joined data, research logs (One Quant Book 7) recording what was tried and what failed, and the know-how of the people who built them. Most of it is protected in only two ways. Either the law treats it as a secret and punishes its taking, or contracts keep the people who know it from using it elsewhere for a while.

Patents and copyright do little of the work. A trading strategy is rarely patentable, and patenting it would publish it; copyright protects the text of code but not the idea it implements, and a competitor who rewrites a model from memory copies no text. So the firm’s protection rests on secrecy, and secrecy is a matter of degree: every departure, every conference talk and every vendor who sees the data leaks some of it.

Three things differ in how fast what leaks loses its value.

• The model’s structure: which effects the strategy trades and how it combines them. It decays slowly, at the pace of the effect itself: crowding and post-publication decay (One Quant Book 7) erode it over years.
• The parameters: the fitted weights, thresholds and sizes. Their value is high but short: they are re-fitted every few months, and a stale copy loses its edge quickly.
• The data: the pipeline that cleans and joins the inputs. Its value decays at the pace of the vendors’ own changes.

The chapter’s model gives each a share of the strategy’s profits that a competitor using it would take, and a half-life, the time for the value of what was taken to halve, which is Book 7’s signal half-life applied to the leaked knowledge.

11.2 Trade secrets and how they are protected

The three conditions are nearly the same in the United States, the European Union and the United Kingdom (Box 11.1). The third is the one a firm controls, and the one litigation turns on: a court asks what the firm did, not what it intended. The reasonable steps are ordinary controls: access on a need-to-know basis, logging of copies and uploads, segregation of duties (One Quant Book 6) between those who write code and those who can move it off the firm’s systems, and contracts.

The criminal case of the hook shows why the words matter. The court held that the source code was intangible, so its transfer by upload was not the transport of a stolen “good” under the National Stolen Property Act; and that the trading system, which the bank kept to itself, was not a product produced for or placed in interstate commerce, so the Economic Espionage Act, as then worded, did not reach it. The 2012 amendment replaced those words with a product or service “used in or intended for use in” commerce, which an in-house trading system is. The civil action of 2016 added a federal route for the owner itself (Figure 11.4).

11.3 Non-competes, garden leave and non-solicitation

Trade secret law acts after the fact and must prove what was taken. Contracts act before: they keep the people who know the secret out of the market for a while, whether or not they take anything.

The two restrictions differ in who pays and in how courts see them. Garden leave is paid employment: the employee keeps salary and benefits, and the restriction is part of a notice period the employee agreed to. A non-compete starts after employment ends; some firms pay for it and some jurisdictions require them to, and whether it is enforceable, and for how long, depends on the jurisdiction and on the terms. In the United States a federal ban was announced in 2024 and abandoned in 2025 (Box 11.2). Deferred compensation (chapter 10) works alongside both: an employee who leaves forfeits the unvested balance, and a hiring firm that buys it out usually waits for the restriction to end.

11.3.1 What garden leave protects

Let a strategy earn $P$ a year, with an edge that decays at rate $\lambda$ whatever happens (half-life $h=\ln2/\lambda$). A departing employee could give a competitor something that takes a share $L$ of the strategy’s remaining profits from the day the competitor starts trading it. If the firm keeps the employee out of the market for $T$ years, at a cost of $S$ a year, and discounts at rate $r$, then

The protected value rises quickly while what was taken is still fresh and flattens once it has decayed; the cost rises in a straight line. The restriction is worth its cost only while the knowledge it keeps out of the market is still worth more than the salary that keeps it there.

def loss_if_start(P, L, lam, r, T):
    return L * P * math.exp(-(lam + r) * T) / (lam + r)


def protected(P, L, lam, r, T):
    return L * P * (1 - math.exp(-(lam + r) * T)) / (lam + r)


def leave_cost(S, r, T):
    return S * T if r == 0 else S * (1 - math.exp(-r * T)) / r


def net(P, L, lam, r, S, T):
    return protected(P, L, lam, r, T) - leave_cost(S, r, T)


def best_length(P, L, lam, S):
    """Where the discounted marginal protection L P e^{-(lam+r)T} equals the discounted marginal cost S e^{-rT}."""
    if L * P <= S:
        return 0.0
    return math.log(L * P / S) / lam

The proposition’s length is an economic optimum, not a legal one. Where the knowledge decays slowly, $T^*$ runs to years, far beyond the three to twelve months of the usual restriction; there the restriction protects only a small part of what is at stake, and the rest must come from trade secret law and from the reasonable steps that keep the secret in the first place. Where it decays fast, a short restriction protects most of it.

11.4 Tutorial: pricing garden leave

The three cases split cleanly. Six months of leave protect 54% of what the parameters are worth to a competitor, 32% for the data and only 19% for the structure. Every restriction in the table is worth more than its cost, since the salary is small against a strategy’s profits. The best lengths are three half-lives for the parameters ($LP/S=8$), two for the structure ($LP/S=4$) and 1.42 for the data ($LP/S=2.7$): 1.42, 4.14 and 1.43 years. The fastest-decaying knowledge is the one a restriction protects best, and the slowest, the model’s structure, is the one it protects least: a four-year restriction is neither usual nor, in most places, likely to be enforced.

What to change next. Rerun with seed=12: the fitted half-lives move to 23.9, 6.6 and 11.7 months and the best lengths to 3.99, 1.65 and 1.38 years, so the ranking is robust and the digits are not. Then let the competitor need six months to rebuild a trading system: the restriction then protects the parameters for longer than it lasts.

11.5 The public litigation record

The disputes that reach a court are few, and most end without a judgement on the merits. What the public record shows is the shape of the claims and how they end (Figure 11.4).

In April 2024 Jane Street Group sued Millennium Management and two former employees in the Southern District of New York (No. 1:24-cv-02783). The complaint alleged breach of contract, tortious interference, trade secret misappropriation, unjust enrichment and unfair competition: that the two employees, who had resigned in February 2024 to join Millennium, had taken a proprietary trading strategy with them. On 6 December 2024 the parties filed a stipulation dismissing all claims and defences with prejudice, each party bearing its own costs and fees. The docket records the allegations and the dismissal; it records no finding that anything was taken.

11.6 Build: the garden-leave calculator

11.7 Exercises

11.8 Problem: The Six-Month Question

11.9 Interview questions