Trading Your Own Money

A long/short book built so that the market's direction does not matter, target beta zero, keep only the relative call. Neutrality raises the Sharpe ratio but shrinks the raw return, which forces leverage, which is where the risk quietly comes back in.

The most common hedge fund structure, own the names you like, short the names you don't, and keep some deliberate net long exposure. The shorts fund the longs and dampen the drawdowns, but they also add borrow costs, gap risk and a beta you almost certainly measured wrong.

The standard test of whether a signal predicts returns, rank the universe, chop it into ten buckets, hold each for a month, and look at the top-minus-bottom spread. It is nonparametric, outlier-proof and produces a tradeable portfolio, which is also why it flatters bad signals.

Rank stocks on any characteristic and you usually end up long one sector and short another by accident. Demeaning the score inside each sector converts an industry bet back into a stock-selection bet, at the cost of some raw return, and usually a big gain in Sharpe.

Exchanges publish, in the minutes before the close, exactly how many more shares want to buy than sell at the current indicated price. That published imbalance predicts which way the closing print will move, and traders who read it fast enough can front-run the auction itself, a strategy that only works because the information really is public but still moves the market.

A stock closes at one price and opens the next morning meaningfully higher or lower, news arrived while the market was shut. Whether that gap tends to keep moving in the same direction or snap back depends on what caused it, and mixing up the two is the fastest way to lose money on this trade.

A documented U.S. equity pattern: the market's return in the first 30 minutes of trading is positively correlated with its return in the last 30 minutes, strongly enough to trade, until you account for what it actually costs to hold a position all day to capture it.

A tail hedge pays off in a crash and loses money almost every other year, so the real design question is never whether it will pay off, it's how much steady bleed a portfolio can tolerate for insurance it hopes never to use.

On February 5, 2018, a fund that let retail investors bet against volatility lost effectively all its value in a single session, and the mechanism that killed it, a forced rebalance that had to buy into the exact spike it was hedging against, spread the damage across the whole VIX futures market.

Selling options or volatility-linked instruments harvests a real, persistent premium, implied volatility tends to run above the volatility that actually shows up. It also means collecting small, steady profits most of the time in exchange for occasional losses large enough to erase years of gains at once.

Options priced ahead of an earnings report tell you exactly how big a move the market expects overnight. Compare that number against what the stock actually does, again and again, and you can find a systematic bias, but the classic trade of selling the straddle assumes the bias holds, and it doesn't always.

Why obvious mispricings survive. Arbitrage is not free money executed by an abstract market, it is a levered trade run by a fund that borrows shares, posts margin and answers to investors, and every one of those frictions can force it out of a position that was right.

A multi-strategy book's Sharpe ratio depends less on how good any single strategy is than on how correlated its returns are with the others, two mediocre, uncorrelated strategies combine into something better than one great strategy run alone, and multi-manager platforms are built entirely around that arithmetic.

Two ways to decide a trade, a rule written in advance and executed without exception, or a human choosing position by position. The difference is not the software, it is whether the book still trades when the PM is on holiday, and it decides whether your edge comes from being right or from being right slightly, often.

Buy spot, short the perpetual future, and collect the funding payment that longs make to shorts. The position has no price exposure, so the yield looks free, until the two legs sit on different venues and one of them liquidates you.

Supply liquidity to an automated market maker, short the resulting token exposure with a perpetual future, and you are left with a clean bet, swap fees against the money arbitrageurs take out of the pool. Hedging removes the direction but not the loss.

Searchers are bots that find profitable transaction orderings, arbitrage, liquidations, sandwiches, and bid for the right to have them included. Finding the opportunity is the easy half; the auction that follows takes most of the money.

Nobody wakes up with a profitable signal fully formed, real ideas come from a handful of repeatable sources (economic mechanisms, market microstructure, other people's published research, and plain observation of what breaks), and knowing which source you're drawing from tells you how to test the idea.

Once a desk has several signals, the question shifts from 'is each one good' to 'how should they be weighted together', and the honest answer sits between naive equal-weighting and a fully optimized regression, because the optimizer's extra precision is bought with extra noise from a short, overlapping history.

A new signal's raw information coefficient is close to meaningless on its own, what matters is how much of its predictive power survives after you control for every signal already in the book, because a signal that duplicates an existing one adds cost and complexity without adding return.

Instead of hedging a stock against an index someone else picked, build the hedge from the data itself: the dominant modes of covariation across thousands of stocks, extracted by PCA, are tradeable baskets in their own right, and what's left over after subtracting them is what you trade.

The whole trading rule is three numbers, where you get in, where you get out, where you give up. Simulating a known mean-reverting spread shows the folklore entry at two standard deviations is far too wide, and explains why desks use it anyway.

The working model behind almost every spread trade, a quantity pulled back toward a level, with the pull proportional to how far it has strayed. Three parameters fall out of a single regression, and they tell you your holding period, your position size and whether the trade survives costs.

Before you can trade a spread you have to build one, and the hedge ratio you pick decides what the residual actually contains. Regressing A on B and regressing B on A give answers 18% apart on the same data, and the gap between them is bigger than any sampling error you will ever quote.

Every row of market data carries a time, but it is usually the time the event happened, not the time you could have known about it. Indexing a backtest by the wrong one is the single most productive way to invent an edge that does not exist.

A backtest built as a loop over timestamped events, where the strategy only ever sees what had already arrived. It is slower to write than a vectorised backtest and far harder to cheat with, which is the entire point.

A normal backtest trades an unlimited amount at the printed price, so its Sharpe is the Sharpe at zero dollars of capital. Capacity-constrained backtesting re-runs the strategy at real book sizes and reports the curve instead of the number.

An options backtest built on end-of-day mid prices can show a beautiful Sharpe ratio for a strategy that would have lost money on every single real fill. Options need their own backtesting discipline because the quoted price is rarely the tradeable one.

The rule your simulator uses to decide whether an order filled, and at what price, usually matters more than the signal. Four defensible fill rules on the same strategy can produce four completely different businesses.

Splitting history into a part you fit on and a part you test on only works if the test part is genuinely untouched. The protocol is about counting how many times you looked, because a hold-out you have peeked at twenty times is no longer a hold-out.

A procedure that splits history into blocks, tries every balanced train/test combination, and asks how often the variant that won in-sample lands below the median out-of-sample. The answer is the probability of backtest overfitting.

Hundreds of published stock-return predictors have been re-tested by independent teams. Most of them shrink badly and many vanish entirely. This page explains why, and how to read a factor paper without inheriting its optimism.

Splits, dividends, spin-offs and rights issues move a stock's printed price without changing what a holder owns. Research that reads raw prices treats those moves as returns, and the errors are large, one-sided and easy to mistake for a signal.

In finance a missing value is rarely an accident. It is usually a halt, a delisting, a failure to report or a stock nobody wanted to trade. Forward-filling deletes that message and replaces it with a fake calm that optimisers find irresistible.

An impossible print and a genuine crash look identical in a histogram. Deleting the first is data hygiene, deleting the second is deleting the market. Most cleaning pipelines cannot tell them apart, and pay for it in both directions.

A point-in-time database remembers what it used to say, not just what is true now. Without that memory, every backtest quietly uses corrected figures that nobody had at the time, and the corrections are concentrated exactly where the money is.

Two datasets with sensible-looking time columns can be merged in one line of code, and that line is where cross-market backtests invent their edge. Daylight saving alone hands a naive join four weeks of free look-ahead a year.

Turning a daily Sharpe ratio into an annual one means picking a day count and assuming returns are independent. Both choices are usually wrong by enough to flip a manager ranking.

A single Sharpe ratio from a backtest is a point estimate from a noisy, finite sample. Bootstrapping resamples the return history to show how wide the range of plausible true Sharpe ratios actually is.

A single Sharpe number hides how a strategy's return changes as you size it up. A capacity-adjusted return curve reports performance as a function of book size instead of pretending there is only one.

A strategy's raw return can look like skill and be nothing more than a tilted bet on well-known factors. Regressing returns on a factor model separates the alpha you can claim credit for from the beta you were carrying for free.

A manager can post a genuinely good return and still watch an investor lose money in the same account, because the investor's own timing of deposits and withdrawals gets baked into one of the two standard ways to measure performance and stripped out of the other.

Instead of asking whether a strategy passes its own backtest, adversarial design asks a second researcher to actively try to break it, and rewards them for succeeding.

Deciding how a strategy will be judged after you've already seen how it performs is not validation, it's rationalisation. The protocol has to be written down before the first backtest runs.

Rebuilding a strategy from the written spec alone, without looking at the original code, catches the bugs and hidden assumptions that reviewing the original code never will.

A backtest's best parameter setting can sit on a broad plateau of nearby settings that all work, or on a narrow spike surrounded by settings that fail. Only the plateau is a real signal.

A backtest that passes every statistical test but can't explain why the market would pay for the signal is still probably noise. Requiring a rationale before the data mining starts filters most of it out for free.

The quoted spread is what you see; the effective spread is what a trade actually paid; the realised spread is what the liquidity provider actually kept once the price moved on afterward. The gap between effective and realised spread is a direct measure of price impact.

When a large order moves the price, part of that move sticks and part of it fades. Splitting the two tells you how much you paid for speed versus how much information you leaked.

Before an order is sent, a desk estimates what it will cost to trade, spread, impact, and the risk of prices moving while you wait, and uses that estimate to pick how fast to trade.

The same trade can be scored as a 22 bps loss or a 40 bps win depending on what you compare it to. Picking the benchmark is not a reporting decision, it is an instruction to the algo about what to optimise.

Every child order you send is evidence. If your slices are the same size, on the same venue, at the same cadence, other people work out what you are doing and the price moves before you get there. Leakage is the part of your cost that never comes back.

Every order is a trade-off between paying the spread now for a guaranteed fill (aggressive) and waiting in the queue for a better price that might never come (passive). The right choice depends on urgency and how likely you are to actually get filled.

You cannot backtest an execution algorithm the way you backtest a signal, your own order changes the very book you're trading against. Simulating execution means modelling queue position, fill probability, and impact, not just replaying historical prices.

Trading depresses the price, and then the book heals, new liquidity slowly refills the levels you just ate. Obizhaeva-Wang models that healing explicitly, which is what lets it say something Almgren-Chriss can't: trading fast and then pausing is different from trading slowly and steadily.

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.

The diversification you counted on when you built the portfolio is measured in calm markets, and calm-market correlations are not the correlations that show up when everything actually goes wrong at once.

Instead of assuming a shape for tomorrow's losses, replay the last few hundred real market days through today's positions, sort the results, and read off the loss you exceed only rarely. Simple, popular, and quietly full of assumptions.

A trading book holds thousands of different instruments, but their risk comes from a few dozen shared drivers. Mapping is the translation step that rewrites every position as a set of sensitivities to those drivers, so the whole book can be aggregated and measured.

A VaR model makes a promise you can check: losses should breach the number on 1% of days and no more. Backtesting counts the breaches, and Kupiec's test decides whether the gap between promised and observed is ordinary bad luck or a broken model.

A precise way to answer "how much of the portfolio's total risk does this one position actually cause," using a mathematical property that guarantees the pieces add up exactly to the whole.

A statistical theory for the worst days specifically, built from the idea that the shape of extremes doesn't have to match the shape of the everyday, and that you can estimate a 1-in-1000-day loss from a lot fewer than 1000 bad days.

A Sharpe ratio computed from a track record is an estimate, and estimates have error bars. For daily data the rule is almost embarrassingly simple, the standard error is about one over the square root of the number of years.