New in a Quant Job

The hour before the bell is where most avoidable losses are prevented. A written pre-open checklist turns "did I remember?" into "did I tick it?", and its job is to catch the problems that are already sitting in your book before you place a single order.

A trading day is not one long stretch of the same activity. Liquidity, volatility and the quality of your signals change hour by hour in a pattern reliable enough to plan around, and most of the discipline of running a book is doing the thing the current hour is actually good for.

A corporate action rewrites the terms of something you already own, without you trading. Share counts, prices, resting orders, hedge ratios and price history all have to be rewritten with it, in the same way, at the same time, or your book and your systems quietly stop agreeing.

A live risk screen shows you forty numbers and you have six and a half hours. This is which five actually predict trouble, how to read them in units of a normal day, and how to tell a bad hour apart from a broken book while there is still time to do something.

A risk limit only works if you treat it as a wall you plan around, not a line you find out you've crossed after the fact.

A great thesis that everyone else already owns can lose money for reasons that have nothing to do with whether the thesis is right.

Cutting size automatically as losses deepen, on a pre-agreed schedule, takes the decision away from the version of you that's already down money and least trusted to make it.

The riskiest moment on a book isn't a single bad trade, it's discovering that several trades you sized as unrelated were actually the same bet wearing different tickers.

A bad P&L day needs a real explanation traced to specific positions and causes, not a story, and not silence, because the explanation is what tells you whether to change anything tomorrow.

A P&L number by itself tells you almost nothing; reading it well means breaking it into pieces you can each check against what you expected.

Most day-to-day P&L wobble is noise your risk model already priced in; the discipline is knowing how much wobble is normal so you notice when a run of days is actually telling you something.

Attribution is the discipline of tracing a P&L number back to the specific position, price move or trade that produced it, before you decide what the number means.

Buying your target size in tranches on a schedule you wrote before you started. It buys you a better average price and a look at how the market absorbs you, and it costs you real money on the trades that work immediately.

Sizing is not one calculation. It is four independent ceilings, risk budget, liquidity, house limits and conviction, and the size you take is the smallest of them, never the average and never the biggest.

Getting out of a big position is a separate skill from picking it: the goal shifts from being right to leaving as little money as possible on the table while you leave.

Once you know what to hedge, you still have to pick the instrument, and the right one trades off tracking accuracy against cost, liquidity and how quickly you can put it on and take it off.

When nothing trades directly against your risk, you hedge with the closest liquid substitute you can find, and the whole skill is being honest about how imperfect that substitute is.

Hedging every risk you can name is not the goal, it's identifying which risks are unrewarded and unwanted, because a hedge you don't need is just a second position paying away your edge.

A stop only protects you if it's executed the moment it's hit, every reason to wait 'just a little longer' is the same reason stops exist in the first place.

Buying more of a position that's already lost money is sometimes the correct trade and sometimes the classic way to turn a manageable loss into a career-ending one, the difference is entirely in whether anything has actually changed since you sized it.

Passing on a trade is a decision like any other, with its own cost if you get it wrong, the discipline is making it deliberately instead of by default.

A systematic signal is built on the data it was trained or tuned on, and there are specific, nameable situations it wasn't built for, knowing which ones justify an override, and logging every time you use one, is what keeps discretion from becoming an excuse.

A good decision can lose money and a bad decision can make money, judging trades only by whether they worked teaches you the wrong lessons.

Writing down the thesis, size, and stop before a trade, not after, is the only reliable way to find out later whether your reasoning was actually good.

The trade that undoes a month usually isn't the first loss, it's the oversized, unplanned trade taken right after it, trying to get the money back immediately.

Holding a position through an earnings report or a central bank decision isn't a passive default, it's an active bet on the event itself, and it should be sized like one.

The hardest part of a regime change isn't knowing markets shift, it's admitting, while it's happening, that the strategy that's worked for months has just stopped.

When realized volatility jumps, every position on the book is suddenly bigger in risk terms than it was yesterday, even though nobody traded a share.

The size you could exit at yesterday and the size you can actually exit at today are different questions, and the gap between them is exactly what shows up on the worst days.

Most research projects fail before a line of code is written, because the question was never answerable. How to turn a vague prompt into a sentence you can be wrong about, bounded in universe, horizon and time, with a stated cost of being wrong.

A hundred ideas go in the top and one gets capital. The stages between are not bureaucracy, they are an ordering that puts the cheapest possible kill first, so the expensive work only ever happens on survivors.

Two ways to start a project, reason your way to a prediction and then test it, or search the data until something lights up. Both find signals; only one of them tells you how much to believe what it found. The difference is entirely about how many things you looked at before the one that worked.

Killing ideas quickly is the highest-leverage skill on a research desk, and the hardest to learn, because by the time an idea is dying you are invested in it. How to tell a dead idea from a badly built one, and how to pre-commit to the decision while you are still neutral.

The researcher owns the evidence; the PM owns the risk and the capital. That split explains almost every disagreement between the two seats, including why a PM will reject a Sharpe 1.8 signal and take a Sharpe 0.6 one.

A research note exists so someone who wasn't in the room can act on a result without re-deriving it, which means leading with the answer and the decision it implies, not with the six weeks of work that produced it.

A backtest chart is a persuasive tool whether or not that was the intent, and the same honest numbers can be made to look far better or far worse depending on scale, start date and what's left off the axis, so the discipline is choosing the chart before you've seen how flattering it is.

A portfolio manager doesn't want to hear how a signal was built, they want to know what it will do to the book they already run, which means answering questions about correlation, capacity and drawdown before anyone asks them.

A kill switch only works if its trigger is a number decided in advance, checked automatically, not a judgment call made by someone watching a P&L line drop in real time.

A backtest is a photograph of history. A monitoring stack is the live camera that tells you, today, whether the strategy behind it is still doing what the photograph promised, before the drawdown does the telling for you.

The gap between the price you saw and the price you got is one number on a report, but it is made of several different mistakes stacked together. Slippage attribution splits it back into the pieces so you know which one to fix.

Every signal has a shelf life, and the backtest that proved it worked cannot tell you when that life has ended. Decay monitoring is the ongoing measurement that catches a signal going stale before the drawdown announces it.

A research environment is the plumbing that stops good ideas dying of friction, a shared data layer, a signal library, a backtest engine that everyone trusts, and a way to compare a new idea against everything that came before it.

A benchmarking harness is the one battery of tests every new signal has to pass through, run the same way every time, so results from different researchers and different years are actually comparable instead of each being a bespoke, unrepeatable analysis.

An alpha library is the desk's memory, every signal ever tested, live or dead, stored with its parameters and evaluation history so nobody re-discovers a failed idea, and so a new signal gets judged against what already exists rather than in isolation.

A research desk that ran 300 backtests last quarter and kept the results in a folder of spreadsheets named "final_v3" has no idea how many of those 300 it takes to find one good Sharpe by chance alone.

Comparing live P&L to what the backtest predicted, trade by trade, is the only way to catch a simulation that was quietly wrong before the gap becomes a pattern nobody can explain.

If nobody, including the original author, can regenerate last quarter's backtest number from scratch, the number can't be trusted, no matter how good it looked at the time.

A strategy moving from research to paper trading to live capital needs a written bar to clear at each step, decided before anyone is emotionally invested in a specific number.

Tick data records every individual trade or quote update as it happens; bar data compresses a time window into open/high/low/close summary numbers. The choice between them is a bandwidth-versus-information tradeoff that shapes what a strategy can even see.

Columnar formats like Parquet store all values of one column together instead of one row at a time, which makes 'scan a billion timestamps and one price column' fast and compressible in a way row-oriented storage never can be.

Splits, dividends and other corporate actions create artificial jumps in a raw price series. A price feed engineer applies backward-looking adjustment factors so that a return computed across the jump reflects what an investor actually earned, not a data artifact.

A point-in-time database can answer not just 'what is true now' but 'what did we believe was true as of last Tuesday,' which is the only honest way to backtest a strategy that would have used data as it looked in the past, restatements and all.

A unit test checks one small piece of code in isolation against a known expected result, run automatically every time the code changes, so a broken calculation is caught in seconds rather than discovered live weeks later.

Git tracks every change to a codebase as a chain of snapshots, so a research or trading team can work in parallel, undo mistakes, and know exactly which version of the code produced any given backtest or trade.

A kill switch is a mechanism, ideally simple and independent of the strategy's own logic, that halts trading the instant something looks wrong, because a bug in a runaway algorithm doing the wrong thing quickly is far more dangerous than one doing nothing at all.

Instead of checking a function against a handful of examples you thought of, property-based testing generates hundreds of random inputs and checks that a general rule always holds, which finds the edge cases a human wouldn't think to write by hand.