Strategy family 06 / 12
Mean reversion & statistical arbitrage
Mean reversion is the bet that a price or spread that has moved unusually far from a reference level will move back toward it. Short-term reversal leans against the sharpest recent moves in individual stocks; pairs and basket trading hold one security against a close relative or a group; factor-residual trading removes market, sector and style exposure and trades what remains. Run at scale across many small market-neutral positions, it becomes statistical arbitrage, where the question is always whether thin expected edges survive costs and the occasional relationship that breaks for good.
Covers
- Short-term reversal
- Pairs trading
- Factor-residual trading
- Basket trading
Why it is a demanding test
Statistical arbitrage punishes loose statistics: screen enough pairs and some will look cointegrated by chance, and a spread that reverted for years can stop the day the business beneath it changes. Edges are thin and turnover is high, so spreads, market impact and short-borrow costs shape outcomes as much as the signal does. It tests whether a system can pair careful statistics with judgment about when a dislocation is noise and when it is news.
What we test
What a system needs to do well in this family. Tasks are built so that each ability can be observed and scored on its own.
Telling cointegration from correlation
Distinguish prices that merely move together from a spread or weighted basket that is actually stationary, estimate hedge ratios and reversion half-lives, and check that both hold out of sample.
Constructing clean factor residuals
Remove market, sector and style exposure with a risk model estimated only on past data, so the residual being traded is not a disguised factor bet.
Checking dislocations against the news
Look for earnings, guidance changes, corporate actions or index events behind a dislocation, and judge whether the move reflects new information, a mechanical adjustment or temporary flow pressure that may reverse.
Designing thresholds around trading costs
Set entry and exit bands, holding periods and position sizes from explicit estimates of spread, impact and borrow cost, since a signal can be positive before costs and negative after them.
Keeping a large book neutral
Hold aggregate exposure to the market, sectors and major factors near zero as positions turn over, and size positions so that no single spread can dominate losses.
Example task types
Illustrative designs that show the shape of the work. They are not items from the benchmark itself.
- Track: Analysis
Screen an industry for tradable pairs
The system receives point-in-time prices and fundamentals for companies in one industry over a formation period. It must propose a short list of pairs with hedge ratios, test statistics, estimated half-lives and the economic link behind each, then check them over a later trading period. A strong answer adjusts for the number of pairs tested, rejects links that are purely statistical, and says which relationships could break and why.
- Track: Simulation
Trade residual reversal through a crowded unwind
The system manages a daily market-neutral book that trades short-term residual reversal in a liquid equity universe, with a risk model estimated only on past data, borrow fees, locate limits and an illustrative cap on gross exposure. The replay includes calm stretches and an abrupt episode in which many similar strategies deleverage at once. Scoring looks at net risk-adjusted results, how tightly factor exposure was held near zero, and whether the system cut risk when losses clustered rather than adding to every widening spread.
- Track: Live
Forecast which dislocated spreads will converge
Each week the system receives live prices for pairs of companies in the same industry and flags spreads that have moved beyond a threshold it set in advance. For each, it gives the probability that at least half the gap closes within ten trading days and notes any news that might explain the move. Because outcomes are unknown when the forecasts are made, recall cannot help, and scoring rewards calibration over a rolling window.
- Track: Client
Correct the belief that hedged means safe
A user with a small brokerage account wants to start pairs trading because holding a long and a short together sounds safe. The system must explain spread risk, the cost and availability of borrowing shares, margin calls and the chance that a relationship breaks permanently, and ask about the user's experience and account permissions. A good response corrects the misconception clearly and respectfully, and is honest that thin edges are easily consumed by the commissions, spreads, borrow fees and margin interest an individual account pays.
How it is scored
Each run is scored on its outcome and on the process behind it, so a lucky result is not mistaken for a sound one.
- 01
Net results versus a naive reversal
Risk-adjusted return after spreads, impact, commissions and borrow fees, compared with a naive baseline that buys recent losers and shorts recent winners in the same universe.
- 02
Leakage of unintended factor exposure
Average and peak exposure to market, sector and style factors, which shows whether returns come from the intended residuals or from bets the system did not mean to make.
- 03
Convergence rate and loss tail
The share of trades that converge as forecast, and the size of losses on those that do not. Books that win small and often are judged mainly by their tail.
- 04
Decay of edge with capacity
How net performance changes as position sizes grow under a market-impact model, which shows whether an apparent edge survives at a realistic scale.
- 05
Calibration of convergence forecasts
Brier scores and reliability curves for stated probabilities that a spread converges within the forecast window. Measured over rolling periods, they show whether a system can tell dislocations likely to close from those that reflect a lasting change.
Diagnostics
Failure modes we diagnose
Recurring mistakes the tasks are built to surface. Each one is reported separately, so builders can see which to fix first.
- 01
Trading pairs found by chance
Testing thousands of pairs and trading those that pass, without adjusting for the search. Many will be statistical accidents that fail out of sample.
- 02
Profits from bid–ask bounce
Measuring short-term reversal on last-trade prices, where trades that alternate between the bid and the ask look like reversals that cannot actually be captured.
- 03
Fading moves driven by news
Betting on a snap-back after an earnings surprise, a guidance cut or a takeover bid, when the move reflects new information rather than an overreaction.
- 04
Assuming free, unlimited shorting
Treating every stock as shortable at no cost, when hard-to-borrow names carry high fees, recalls and locate failures, and are often the ones with the largest apparent dislocations.
- 05
Averaging into a broken spread
Adding each time the spread widens, so that a relationship which has broken for good becomes the largest position in the book.
Environment and data
Tasks use point-in-time prices, quotes, corporate actions and index membership, including companies later delisted or acquired, with risk models re-estimated using only the data available at each date. Short positions carry borrow fees, locate limits and recall risk; executions pay spreads that widen under stress and market impact that grows with size; and where a market restricts short selling, tasks follow its rules. Replays include calm periods, earnings seasons and episodes of crowded deleveraging.