In a bookshop, a new release sits beside a work publishers have kept reprinting for generations. You know something extra about the latter: interest in it has survived changing tastes and competing titles. Lindy Effect describes a tendency for a longer past, under certain conditions, to suggest a longer estimated time still to come. Age has no magical power here. Survival gives us information. Bitcoin’s operating history can therefore matter without guaranteeing a new service’s safety or a future price. To use the idea well, first say what you are watching and what would count as its end.
In his 2023 paper, Toby Ord distinguishes a broad tendency from precise mathematical rules. A book makes the distinction tangible: are you tracking whether the work remains in print or when a particular copy falls apart? Those are different questions. A technological principle may likewise persist while individual products and versions disappear. One category’s history does not automatically carry over to another. Define the beginning and end of the life you are studying before drawing conclusions from its age. A project might end when development stops, users leave or operations cease; each choice produces different data. [1]
Nassim Nicholas Taleb illustrates the idea with a power-law model, specifically a Pareto model. It assigns probabilities to different lifespans and allows for very long ones. Its parameter α > 1 controls the shape of the lifespan distribution; t is the age already reached. Conditional on surviving to that age, the model’s mean total lifespan is αt/(α − 1), whereas its mean remaining lifespan is t/(α − 1). With α = 2 and t = 10 years, that gives another 10 years, or 20 altogether. Past and future years must not be confused. This is a result of the chosen model, not a promised end date for a particular book or technology. The parameter needs justification too; a pleasing answer is not a reason to choose it. [2]
Ord points out that the mean and median answer different questions. The mean takes all durations into account, so unusually long lives can pull it up considerably. The median divides the possibilities into equally probable shorter and longer halves. In the same Pareto model, with α = 2 and an age of 10 years, the median remaining time is 10 × (√2 − 1), about 4.14 years, not 10. The square-root symbol simply expresses the calculation for this model. At α = 1, the remaining median is 10 years, while the mean is not finite. That is not immortality: an infinite lifespan has zero probability in this model. Check what a quoted number means before using it. [1]
Imagine a process whose known hazard rate stays constant. The handbook of the US National Institute of Standards and Technology (NIST) describes this with an exponential distribution. The rate, written λ, describes how quickly endings occur among cases still surviving; it is not simply a percentage of the original group. At λ = 0.1 per year, the mean remaining lifespan is 1/λ = 10 years, whether the case has survived 5 or 20 years. The conditional total mean is therefore 15 or 30 years, respectively, but the remaining mean has not changed. This counterexample matters: Lindy Effect is not a rule to apply to everything old without checking. [3]
For real projects, we usually do not know the risk in advance. Ord shows how Lindy Effect can emerge from a mixture of cases with different resilience. Less resilient ones disappear, leaving a growing proportion of more resilient cases among survivors. A longer history then gives us information about the group a case belongs to. Nothing inside it need have improved. We cannot automatically conclude that aging makes it stronger or that every shock helps it. Under certain conditions, this selection effect can arise even when each case’s own risk is increasing. [1]
Imagine following 100 projects for 10 years. We have 40 that ended and 60 still active. Do not assign those 60 a final lifespan of 10 years or infinity: we know only that they lasted until at least the end of observation. For each project, this is a lower bound, not its final lifespan. NIST calls this right censoring. Despite the unfamiliar name, it means admitting that the ending is unknown, not deleting inconvenient evidence. We also need the failed projects and their times of failure. Starting with today’s winners alone would hide those that did not pass the same test of time. [4]
Bitcoin.org describes open development, fixes for software bugs and the risk of losing access to funds. Bitcoin’s long history tells us about the network; it is not automatically the history of a newly released wallet, a custodian or modified code. Network survival does not prove a particular custody arrangement is safe. With Self-custody, you manage your own private keys: the secret data used to authorize spending your bitcoin. You still need to understand how to back up and restore that access. Network Effect, the usefulness of other people’s participation, might also help explain persistence, but that is a particular mechanism to investigate. A new attack or dependency can change the conditions under which past experience was gathered. [5]
Lindy Effect is most useful as the beginning of an argument, not its conclusion. Following Ord’s distinctions, clarify the comparison group, the event that counts as ending, data selection and the difference between mean and median. Then ask what still works as before and what has changed. For money, Store of Value — preserving purchasing power into the future — matters too. Continued existence alone does not guarantee it. A longer estimated lifespan proves neither a claim’s truth, future demand nor an asset’s price appreciation. Take history seriously, but keep examining the particular risk. [1]
For the clearest picture, read this entry together with Bitcoin, Network Effect, Self-custody, Store of Value.