The Rational approach to a Thursday night

There is probably no cultural phenomenon as prevalent across Sonipat as Thursday night parties. For as long as I've been at Ashoka, and years before that, Thursday nights have gone from being a time, to a verb (‘thursday nighting’), to a mood, to a state of mind. Recently, however, Thursday nights have become more of a game theory case study than a pureplay enjoymentmaxxing competition. Constant patrolling, suspensions, and a seemingly ubiquitous atmosphere of surveillance. However, if I were to tell you whether or not one should go out on a thursday night to ‘maximise their utility’, how would we go about making that measurement? It’s a really interesting question to ask, and the answer is more optimistic than it might seem.

The Exposition

On a rather beautiful Friday morning in January, I wake up at 1pm and scroll through my phone looking for last night’s pictures. My focus soon shifts, however, from looking for pictures to doomscrolling through every app I could, including GMail. The second email from the top, sent at 3 in the morning by the Dean of Student Affairs is titled, rather ominously, “MGI Again”. That thursday, 1500 students went out, and more than 750 were out post 12am. 15 had BAC levels above 100, and 1 had to be referred to an external hospital. For lack of a better word, that Thursday was a shitshow.

Buried inside that email was a genuinely interesting question: what actually happened to the risk of a Thursday night when the reward — the promise of a record crowd — pulled in three times the usual turnout? And is the trade-off that us, individual students with careers and academics at stake, are making, night after night, actually a good one?

I wanted to find out, so I built a model.

Treating a night out like a trade

Every Thursday night is, structurally, a bet. You put in a few hours and a couple grand, and in exchange you get an uncertain payoff: a good time, some new memories, an outside chance of a much worse outcome. That's the same shape as a financial position — an expected return, and a variance around it — which is why the instinct to reach for a Sharpe ratio isn't as much of a stretch as it sounds.

But a Sharpe ratio assumes our risk is symmetric: it penalizes a great night exactly as much as a bad one, because it only measures how spread-out the outcomes are, not which direction they're spread in. That's a bad assumption here. The overwhelming majority of Thursdays end in nothing worse than a mild hangover. A small minority end in a breathalyzer test and a suspension. And within that tail, a much smaller slice cascades into something genuinely severe — barred from campus placements, or repeating a semester entirely. That's a heavily lopsided, "fat left tail" distribution, the same shape you'd get from writing insurance or selling options. The right tool isn't Sharpe. It's something that only penalizes the downside — a Sortino ratio — combined with an honest accounting of what that downside actually costs.

What the record night revealed about risk

The email gave me real numbers to anchor a risk curve: 15 flagged incidents out of roughly 1,500 people out, against 2–3 flagged incidents on a typical night of perhaps 500. Fitting a curve through those two points shows something that matters more than either number alone: total incidents scale faster than turnout does. Specifically, incidents grow roughly with N^1.63 — so a 3x jump in turnout produced a 6x jump in flagged cases, and the per-student probability of getting flagged roughly doubled, from about 0.5% on a typical night to about 1.0% on the record one.

That's not what you'd expect if everyone's individual drinking decisions were independent of each other. It's what you'd expect if the systems meant to catch problems early — wardens, the infirmary, gate staff — have fixed capacity. Once 750 people miss curfew and 250 are still outside at 12:30, those systems stop intervening early and start only catching the worst cases, the ones bad enough to need a hospital referral. The crowd didn't just get bigger. It got riskier per person, because the safety net got thinner per person.

Putting a rupee figure on the tail

Here's where the piece gets less comfortable. "Risk" is easy to gesture at in the abstract; it's harder to sit with once it has a rupee figure attached. So I built three severity tiers, each based on real costs rather than a made-up scale:

• Low (₹2,000): a bad night with no suspension — a rough estimate of a medical referral plus an on-the-spot legal cost, if you're unlucky enough to draw police attention without it escalating further.

• Medium (₹12,000): the above, plus the administrative and reputational cost of a single one-week suspension that doesn't yet touch placements or your degree timeline.

• High (₹19.2 lakh): the full cascade — one extra semester of tuition and residence fees, roughly six months of delayed entry into the job market valued at the average on-campus placement package (₹11.6 lakh, per Ashoka's 2025 placement report), plus the placement opportunity itself, discounted for what you could still recover through an off-cycle search.

That last number is worth sitting with. It is not a fine. It is not a slap on the wrist. It's a figure large enough to represent a genuine financial event in a young person's life — and it exists as a real possibility every single Thursday, at a probability that's small but not negligible, and rising with the size of the crowd.

Running the numbers

Combining the risk curve with those three costs, and a reward function that treats a typical night's own spend (₹2,000, on entry, drinks and transport) as a lower bound on its value — boosted modestly by network effects as more people turn out — produces expected values for a record-sized Thursday (1,500 people out) of:

Severity scenario	Expected value	Sortino ratio
Low	+₹2,614	1.60
Medium	+₹2,514	0.37
High	−₹16,581	−0.10

Under the Low and Medium framings, going out on a big night is a perfectly good bet — mildly positive expected value, with a healthy Sortino ratio in the Low case. Under the High framing, it flips to solidly negative. The three scenarios don't disagree about whether risk exists. They disagree about how close you personally sit to the trapdoor.

Solving for the turnout size that would actually maximize expected value tells the same story from a different angle. Under Low severity, the "optimal" crowd is around 189,000 people — in other words, there's effectively no natural ceiling; go as big as you like. Under Medium, it's about 29,500 — again, no realistic constraint. Under High severity, priced honestly, the number collapses to about 13 students.

That is the actual finding of this model, and it's a strange one to sit with: a rational actor who has fully priced in the cascading cost of a repeated semester and a lost placement should, by pure expected-value logic, basically stay in. And yet on the night in question, roughly 1,500 people didn't. Something other than a personal cost-benefit calculation was doing the deciding.

Why 1,500 people made the same "mistake" at once

This is where the ResLife email's own diagnosis is more useful than it probably intended to be. It attributes the record night not to a shift in student character, but to a specific policy change — a recent relaxation of Main Gate Infraction penalties. More than a demand-curve shift, It's a shift in what everyone believed everyone else was about to do.

Sociologist Mark Granovetter's threshold model of collective behavior, originally built to explain riots, can be used to solve exactly this. Each person has a threshold — the number of other people they need to see already committed before they'll join in. Attendance isn't the sum of independent decisions; it's a fixed point of everyone's beliefs about everyone else's beliefs, which is the same logic Keynes used to describe a beauty contest, and the same logic behind a bank run: a bank is fine as long as everyone believes it's fine, and collapses the moment enough people believe others are about to pull their money out.

Modeling Ashoka's student population this way — a small minority of "early joiners" who go out regardless, and a majority who only go once a real crowd is visibly forming — produces exactly the kind of behavior the campus saw: two stable resting states, a quiet one and a "lit" one, separated by a tipping point. Nudge the initial buzz from just below that tipping point to just above it, in the simulation, and turnout swings from a couple of dozen students to essentially the entire eligible population. Nothing about any individual's risk tolerance needs to change for that to happen. The system just needs to cross a threshold in collective belief.

That reframes the whole question. A Sharpe or Sortino ratio treats a Thursday night as a bet you price and choose for yourself. Granovetter's lens says the real decision variable isn't how much to drink — it's whether your own attendance nudges the crowd's shared belief past the point where it snowballs into the high-turnout, high-risk equilibrium in the first place. Once that tips, the convex risk curve does the rest, and 1,500 individually reasonable-seeming choices add up to a night the infirmary can't handle.

What this is actually worth doing with

None of this is an argument that going out is irrational, or that a single Thursday night is a bet nobody should take. The honest reading is narrower and, I think, more useful: the "correct" answer depends enormously on where you personally sit on the severity spectrum, and that's information most students never explicitly work out for themselves. Someone with a clean disciplinary record, several semesters of runway, and a placement season still a year off is genuinely closer to the Low scenario. Someone already carrying a warning, in their final placement cycle, or entering exam season is much closer to the High one — and the model says that person's expected value on a crowded Thursday is meaningfully negative, whether or not it feels that way while everyone else is walking out the gate.

The crowd, in other words, doesn't know which of those two students you are. It only knows how many people are already out.

Methodology note: the risk curve is fit directly to the two data points in the original ResLife email (15 flagged incidents at ~1,500 turnout; 2–3 at an assumed typical turnout of 500 — the email doesn't state the latter, so it's the single most important number to correct if a real figure becomes available). The severity costs are built from Ashoka's published 2026–27 fee structure and 2025 placement report, plus several explicitly labeled estimates (a ₹1,500 medical-referral cost, a ₹3,000 legal-cost placeholder, and a ₹10,000 suspension administrative cost) that are best treated as illustrative rather than researched. The cascade simulation is a parametric version of Granovetter's threshold model, calibrated to produce the same order-of-magnitude gap between calm and "lit" nights that the university has observed. A fully editable version of the model, with every assumption exposed on its own sheet, is available on request.