Regime tells you whether the world is risk-on or risk-off. Character tells you whether to fade moves or follow them — the single most useful thing to know before you place a trade. Sharpe and drawdown tell you whether the engine is currently working and how rough the ride has been. Leverage tells you how big to go. Each one is close to useless on its own — they are a set. Every panel below has a ▸ Explanation link, or hit Show explanations above to open them all at once.
Variance-ratio estimate over 252 bars. Below 0.5 the last move tends to be given back (trade reversion, no stop-loss). Above 0.5 moves extend (trade momentum, stop-loss is justified). Inside the band there is no serial structure to trade — the honest answer is "sit out", and the panel is worth publishing precisely because it says that out loud.
| Asset | Hurst | Half-life | Read |
|---|
What it says. Does this market snap back after a move, or keep going? Blue means it snaps back. Orange means it keeps going. Grey means neither.
What to do with it. Pick your tactic from the colour before you pick your entry. Blue: buy dips, sell rips, be patient. Orange: buy strength, cut losers quickly. Fading a trending market and chasing a reverting one are the two commonest ways to lose money slowly.
When to ignore it. When it's grey — that's the panel telling you honestly that there's no pattern to trade. Also note it can't see a slow steady grind higher; it only measures whether one move predicts the next one.
Fit dz = −θ(z−µ)dt + dW and read half-life = ln2/θ. This is the statistically robust way to set a holding period: it uses every bar in the series, not just the handful of bars where a trade actually fired, so it does not inherit the small-sample noise that wrecks momentum holding-period estimates.
| Spread / asset | θ | Half-life | Max hold |
|---|---|---|---|
| GLD − 1.68×GDX | 0.069 | 10.0 d | 20 d |
| SPY intraday z | 0.42 | 1.6 d | 3 d |
| CAD/AUD | 0.031 | 22.4 d | 45 d |
| BTC 4h reversion | 0.11 | 6.3 bars | 13 bars |
| Silver vs miners | 0.018 | 38.5 d | too slow |
What it says. When a price gets stretched away from normal, how long does it usually take to travel halfway back?
What to do with it. Use it as an egg timer. Half-life of 10 days means expect roughly 10 days to collect half your move, and don't plan to sit there much past 20. It stops "it'll come back eventually" from turning into a six-month hostage situation.
When to ignore it. When the panel on the left says that market is trending — if prices aren't coming back, there's no "back" to time. And "too slow" doesn't mean the pattern is fake; it means your money is better used elsewhere while you wait.
Rolling hedge ratio, not full-sample — a full-sample regression is a look-ahead bias, since you could not have known the coefficient at the time. Enter beyond ±2σ, exit inside ±1σ. Hover for the value on any date.
What it says. Two things that normally travel together have drifted apart. This line measures how far apart, in "how weird is this" units. Zero is normal.
What to do with it. Past +2 or −2, they're unusually far apart — buy the cheap one and sell the expensive one. Back inside the shaded band, the gap has closed: take it off. You're not betting on either one going up. You're betting the gap shuts.
When to ignore it. When the table to the right says that pair failed its test. Two things drifting apart is only a trade if they're actually tied together — otherwise you're just buying something that's falling for a good reason.
Augmented Dickey–Fuller t-statistic on the residual. More negative is more stationary. Critical values −3.90 / −3.34 / −3.04.
| Pair | t-stat | ρ(returns) | Verdict |
|---|---|---|---|
| GLD / GDX | −3.36 | 0.71 | ✓ 95% |
| CAD / AUD | −3.95 | 0.63 | ✓ 99% |
| XLE / CL front | −3.11 | 0.58 | ~ 90% |
| SLV / SIL | −2.88 | 0.79 | ✗ fails |
| KO / PEP | −2.14 | 0.48 | ✗ fails |
KO/PEP is the teaching case and belongs in the Learn section: the two are significantly correlated (ρ=0.48, p=0) and not cointegrated. Correlation is about tomorrow's returns; cointegration is about whether the prices stay tethered over years. Most retail pair traders screen on the wrong one.
What it says. Which pairs are genuinely joined by a rubber band, and which just happen to wiggle in the same direction. The t-stat is a tethered-ness score: more negative means more firmly tied.
What to do with it. Only trade the green ones. Re-check weekly — pairs come apart, and a pair that drops out of green is a pair to close, not to double down on.
When to ignore it. When you're tempted by the correlation column. Coke and Pepsi move together most days and still drift apart forever. Correlation is about tomorrow; this test is about the next few years. Trading the wrong one is the classic pairs mistake.
Annualised Sharpe of a Bollinger reversion rule at four one-way cost assumptions. A reversion rule on ES futures goes from Sharpe 3 to Sharpe −3 on one basis point. The number that matters is the breakeven cost: mean gross return per bar divided by mean turnover per bar. If that is not comfortably above what you actually pay, the strategy does not exist.
What it says. What fees and spread do to a strategy that looks brilliant on paper. Each bar is the same strategy, charged a bit more per trade.
What to do with it. Find where the bars turn from blue to red. That's the cost level at which the edge dies. Compare it with what you actually pay per trade. If your real cost sits past that line, the strategy doesn't exist — it's an artefact of a spreadsheet that forgot to charge you.
When to ignore it. Never. This is the check that kills most ideas, and it's much better that it kills them here than after three months of live trading.
Six checks, run automatically on every published backtest, displayed with the result. This is the highest-leverage thing on the whole list — the site currently publishes a track record and three bots with, by its own wording, "no returns, Sharpe ratios, or drawdown figures". Sharpe and drawdown are the performance report; returns alone are ambiguous until you say what the denominator was.
What it says. Six questions that separate a real backtest from a good-looking accident. A tick means it passed. An exclamation means there's a known weakness.
What to do with it. Read the exclamations first. They rarely mean "bin it" — they mean "trade this smaller than the headline number suggests". The two that matter most: survivorship (the test quietly ignored the companies that went bust, which makes any bargain-hunting rule look better than it was) and recency (it worked in 2019, the question is whether it works now).
When to ignore it. If you trade on judgment rather than a tested rule, none of this applies to you — though it's a fair list of questions to ask anyone selling you a system.
Three numbers, not one: current drawdown, maximum drawdown, and maximum drawdown duration — the longest it took to recover. They rarely happen over the same window. Duration is the one that actually makes people switch a model off, and it is the one nobody publishes.
What it says. How far below its best the account has been, and — the bit almost nobody publishes — how long it stayed down there before making a new high.
What to do with it. Depth tells you whether you can afford it. Duration tells you whether you can stand it. Ask yourself honestly, before you start: would I still be here after that many months of nothing? Most people quit on duration, not depth. A 10% loss that takes a year to heal breaks more traders than a 20% one that recovers in a month.
When to ignore it. In the first few months. There isn't enough history yet for the worst case to have shown up — the number is flattering, not reassuring.
Everything falls out of one equation, g(f) = r + f·m − s²f²/2. Set the derivative to zero and f* = m/s²; at that leverage the growth rate is r + S²/2 — Sharpe squared, which is why a boring high-Sharpe book beats an exciting low-Sharpe one once you can borrow. Read the box on the left; everything to the right of it is the working.
The row worth staring at is growth unlevered 9.80% against a mean return of 11.23%. Nobody charges you that 1.43% — it is what volatility does to compounding. It is also the reason a calmer strategy with a lower headline number can finish ahead of a wilder one, and the reason risk management is a return decision rather than a comfort decision.
What it says. How big your position should be. Everything above the bottom row is the working; the bottom row is the answer, plus which limit is stopping you going bigger.
What to do with it. Use the bottom number. Two ideas underneath it are worth carrying around: swings cost you money even when the average is fine — this book returns 11.2% on average but only actually grows at 9.8%, and the gap is pure volatility. And steady beats spectacular — once you can borrow, a boring consistent edge compounds faster than a lumpy exciting one.
When to ignore it. If you use no leverage at all, this is just telling you your edge is modest enough that sizing isn't your problem. The one line to never ignore is the worst-day row: the formula assumes crashes are rarer than they are, which is why the cap sits next to it.
The finding is asymmetric and worth publishing as a headline: equity calendar effects have largely died, commodity ones have not, because the demand behind them is real rather than speculative. Each row ships with its full per-year P&L and per-year max drawdown, because a trade that fires once a year has almost no sample and the drawdown column is the only honest risk disclosure.
| Trade | Window | Yrs +ve | Worst DD |
|---|---|---|---|
| Long RB gasoline | 13 Apr → 25 Apr | 14/14 | −$9,816 |
| Long NG (June contract) | 25 Feb → 15 Apr | 14/14 | −$7,470 |
| January effect (small caps) | 31 Dec → 31 Jan | 1/3 | decayed |
| Heston–Sadka month-on-month | monthly | — | dead post-2002 |
What it says. Trades that work because of the calendar rather than the chart — people drive more in summer, so petrol demand rises every spring, and the futures price tends to follow.
What to do with it. Read the last two columns together. A trade that won 14 years out of 14 sounds unloseable until you see it went nearly $10,000 underwater mid-trade in one of those years. Size for the hole, not the hit rate.
When to ignore it. The equity ones. Once everyone knew about the January effect, it largely stopped working. Commodity seasonals have held up better because the demand behind them is real people buying real fuel, not traders front-running a pattern.
Principal components of the return covariance matrix are a factor model you can build from price data alone — no expensive fundamental history required. The share of variance explained by the first eigenvector is a clean crowding and breadth reading: when it spikes, everything is one trade and your diversification is imaginary.
| PC1 variance share | 48.2% |
| PC1 share, 1yr ago | 31.7% |
| PCs to reach 80% | 4 (was 9) |
| Reading | crowding |
Pairs naturally with the existing sector heatmap and gives the composite regime score a breadth input it currently lacks. Traded directly as a momentum signal it loses money — the value here is as a risk lens, not a signal.
What it says. How much of the market is really just one trade wearing different hats. When the top number climbs, everything is moving together.
What to do with it. Treat it as a warning light on your own diversification. If you hold eight names and this reads 48%, you don't have eight positions — you have one position and seven copies of it. The fix is smaller size, not more names.
When to ignore it. If you only hold one or two things, this is telling you about the market, not about you. It's also not a buy or sell signal — traded directly it loses money. It's a risk gauge.
The same logic as this page, as TradingView scripts: LL Quant Layer plots Hurst regime, cointegration spread, Kelly console and the cost cliff.
