EBOLIM Study Library

The concept

• A backtest run first misleads: too many degrees of freedom (entry, exit, stop, size, dates), path dependency (each trade blocks capital for others), and first-trade bias (one lucky opening trade colours the whole curve).
• A signal test isolates the edge: find every instance of the raw condition, measure the return over a fixed forward window (week/month/quarter), with no sizing or exits. One honest question — does this condition predict anything?
• Only after the signal passes do you add exits, stops and sizing, then backtest.

• Observation — you notice a repeatable pattern (e.g. oversold dips in an uptrend bounce).
• Research — you check it holds on other names/periods.
• Hypothesis — a testable claim: "RSI(2)<20 above the 200-day precedes a positive 5-day return."
• Experiment — code the rule, find every instance.
• Analysis — win rate, average return, distribution.
• Conclusion — supports or refutes; you can revisit any step.

• Trigger — true only on the event day (the entry). Close crosses above SMA200.
• Filter — true across many days; screens out unwanted signals. above the 200-day, full-history, liquid.
• Value — a number for ranking when many fire the same day and capital is finite (lowest ATR, highest momentum, best backtested expectancy…). Mixing the three types corrupts a test.

• Signal — a rule fired; a potential opportunity, not a trade. Tested for predictive value.
• Strategy — a validated signal dressed with entries, exits, stops, and sizing.
• Model — several strategies combined to cut portfolio volatility and smooth returns; defines when each is on and how capital is split.

• A robust strategy shows a wide plateau of acceptable results across a parameter range; an overfit one shows a narrow peak with sharp cliffs.
• Pick parameters near the centre of the plateau, not the optimisation peak. A rule that works across 30–150-day averages beats one needing exactly 47.
• False precision (an "optimal" 3.78 std-dev) is a red flag, not a feature.

• Base-rate fallacy — 95% confidence ≠ 95% chance the strategy is real. If few ideas are truly good, most "significant" hits are false positives. Demand p<0.01.
• Multiple testing — test 150 ideas at 1% and you're ~78% likely to find a random winner. Correct for the number of trials (Bonferroni / deflated Sharpe).
• Bias–variance — too simple misses the edge, too complex overfits; total error is U-shaped.
• Meta-look-ahead — you already know 2008/2020 happened, so you unconsciously design to dodge them. Useless out of sample.

• Descriptive stats summarise the sample you have (mean, std, skew).
• Inferential stats use that sample to make a claim about the unknowable population — all future outcomes — with a stated confidence, never certainty.
• Statistics describe a sample; parameters describe a population. Sampling variability means a sample can mislead.

• Null (H₀) = "no edge / returns ≤ 0"; alternative (H₁) = "there's an edge". You assume H₀ until evidence forces its rejection.
• p-value = the probability of a result this extreme if H₀ were true. Small p → reject H₀. Standard cutoff 0.05, but for strategy hunting use 0.01 or lower.
• Type I = false positive (adopt a dud). Type II = false negative (discard a gem). In trading, Type I is the expensive one.
• Test a mean: t = (x̄ − μ) / (s/√n) against n−1 d.f.

• Empirical rule: in a normal distribution 68 / 95 / 99.7% of data sit within 1 / 2 / 3 std-devs.
• Skew — asymmetry. Positive = rare big wins (trend-following); negative = rare big losses (selling insurance).
• Kurtosis — tail weight. Markets are leptokurtic (fat tails): 2σ+ moves happen ~3× more often than normal predicts, so std-dev alone understates risk.

• Pearson's r ∈ [−1, +1] measures the strength/direction of a linear relationship, built from returns, not prices.
• r² (coefficient of determination) = the fraction of one variable's variation explained by the other. r = 0.71 → r² ≈ 0.50.
• Correlation ≠ causation — a lurking third variable can drive both.

• Random walk — past points don't forecast the next, and it assumes a flat drift. Too limited to model real prices.
• Martingale — adds a drift term, so it can slope up (equities on balance rise). A better price model.
• Arcsine law — even in a random series, prices spend far more time trending away from a starting point than hovering near it. The least likely outcome is "return to start". A statistical hint that trend-following isn't crazy.

• Sharpe = excess return ÷ total volatility — penalises up- and down-swings equally (its flaw).
• Sortino = excess return ÷ downside deviation only — right when drawdown is what hurts.
• Calmar = annual return ÷ max drawdown; feeds position sizing directly. Practitioner band: ≥2 acceptable, ≥3 ideal.
• Information ratio = (strategy − benchmark) ÷ tracking error — consistency of outperformance.
• Sharpe assumes normal returns; a steady +12%/yr can score worse than a wild +62/−38 pair. Always read it beside drawdown and the distribution.

• Expectancy = (Win% × Avg Win) − (Loss% × Avg Loss). ≤ 0 and no win-rate saves you.
• R-multiple (Tharp) = profit ÷ initial risk (entry − stop); risk-normalises every trade. Target avg > 1R.
• MAE / MFE — worst/best unrealised excursion. MAE tells you where to set stops so you don't get shaken out of eventual winners; MFE shows profit left on the table.
• SQN (System Quality Number, Tharp) grades a system from its average R and trade count — reliable at ≥30 trades.

• Profit factor = gross profit ÷ gross loss. Sane band 1.25–10; most reliable 3–5; >10 usually means curve-fitting.
• Payoff = avg win ÷ avg loss (trend > 1, reversion often < 1); win rate; ≥30 trades floor, 100+ ideal.
• VWR (variability-weighted return) — a refinement that rewards returns earned in calm periods and penalises those earned in volatile ones, fixing part of Sharpe's blindness.

• "I risk $X to my stop" ignores time and volatility. Real risk = potential portfolio change per unit of time (2%/day ≠ 2%/year).
• Mark-to-market: your position is worth today's price — not entry, not the peak. No "playing with house money".
• If you missed an entry, the right size is the full size at the current price — past trades don't change future probabilities.

• Equal dollars ≠ equal risk — a volatile name dominates an equal-weight book. Size inversely to volatility so each position contributes equal risk.
• Clenow's ATR sizing: shares = (risk% × equity) ÷ ATR.
• Simple stop-based sizing: shares = risk$ ÷ (entry − stop).

• Build a correlation matrix from returns. Seek correlations near 0 — strongly negative can erode the winners.
• Row-sum shortcut: to add to a book, sum each candidate's correlations to current holdings and pick the lowest sum.
• Mistakes: over-diversifying (dilutes the edge) and false diversification (things that secretly move together).

• Vol scaling: scale factor = target vol ÷ current vol; hold the rest in cash. Leverage scales risk and return, so under risk-free financing the Sharpe/Calmar are ~unchanged.
• Vol conversion: σ_annual = σ_daily × √252. 2% daily ≈ 31.7% annualised.
• VaR = worst loss at a confidence over a horizon; CVaR = the average loss beyond VaR — the honest tail number 2008 exposed.
• Leverage to a mandate: if a strategy's max DD is 10% and you tolerate 20%, 2× matches it.

• Investors are risk-averse: the pain of a loss is ~2× the pleasure of the same gain (prospect theory).
• Risk appetite falls as the account grows — a 20% loss on $1M bites harder than on $10k.
• Preserve capital first: a −50% loss needs +100% to recover, so protecting the downside beats chasing return.

• Don't fight the trend — trade pullbacks inside it, entering near where trend-followers stop out, betting on the snap-back.
• Clenow's teaching model: bull regime = 40-EMA > 80-EMA; pullback trigger = price 3×ATR below the 20-day high; exit on trend break or after ~20 days.

• Time-series (trend-following) — a name's own past return predicts its next; each asset judged alone. Your GLD breakout, and the century-long trend-following record.
• Cross-sectional (relative strength) — rank names against each other; leaders keep leading. What the Scanner's momentum board does.
• Magnitude matters for time-series; only the ranking matters for cross-sectional.

• Relative strength = price ÷ benchmark. Rising = outperformance — which can mean "down less", not just "up more". Raw RS values are meaningless across names; rank, don't compare levels.
• RRG plots RS-Ratio (relative trend) vs RS-Momentum (its rate of change); names rotate clockwise Leading → Weakening → Lagging → Improving, ~90% of the time.

• VWAP = Σ(price×vol) ÷ Σvol, cumulative from the open. Price above a rising VWAP = buyers in control; below a falling VWAP = sellers.
• The "ultimate sentiment indicator" — it measures real money, not opinion. Intraday volume is U-shaped (heavy open & close, quiet midday).
• Anchored VWAP starts the calc at a meaningful event (earnings, gap, swing high/low) and doesn't reset — cumulative sentiment from that anchor.

• ATR = smoothed True Range (greatest of H−L, |H−prevC|, |L−prevC|) — the industry vol unit for stops and sizing.
• Bollinger = 20-SMA ± 2σ. %b locates price in the band; BandWidth tracks the cycle — a Squeeze (125-day low) precedes expansion. Volatility is "extreme-seeking", not mean-reverting.
• A band tag is not a reversal by itself: ADX>20 (trending) → continuation ("walking the band"); ADX<20 (range) → mean-reversion. Multicollinearity: five momentum indicators aren't five confirmations.
• Keltner uses ATR bands vs Bollinger's std-dev bands — smoother; a Bollinger-inside-Keltner squeeze is a tight-coil signal.

• Classic MACD = 12-EMA − 26-EMA, plus a 9-EMA signal line. Its flaw: the value isn't comparable across names or dates.
• MACD-V divides by ATR, so a "big" reading means the same on a calm stock and a wild one — momentum on a standard scale.

• An oscillator's value is velocity (direction of change); its slope is acceleration. Acceleration turns before velocity — momentum leads price, which is why divergences precede reversals.
• Overbought/oversold are conditions, not signals — in an uptrend, "overbought" is strength, not a sell.
• A valid signal has three steps: condition (a setup forms) → trigger (a specific event) → confirmation (price itself validates). Momentum alone never gets you in.

• The market discounts everything and looks months ahead. Use closing prices (the day's most important).
• Three movements: primary (year+), secondary (weeks–months, the deceptive counter-move), minor (days).
• Confirmation — two averages must agree; volume should expand in the trend's direction.

• Advance-Decline line — net advancers, cumulative; diverging from the index warns of thin participation.
• New highs / new lows — leadership; daily New Lows are an under-used but robust bear-onset signal (Vince & Williams).
• % above the 200-day (breadth), McClellan Oscillator/Summation, Zweig breadth thrust (10-day A/D from <40% to >62%), TRIN/Arms.

• VIX = 30-day implied vol of the S&P, annualised; mean-reverts ~20; inverse to price (~−0.8); spikes at bottoms but peaks before the low.
• Term structure: contango (calm, up-sloping) vs backwardation (stress, down-sloping); a flip to backwardation is an early-warning.
• VVIX / SKEW / MOVE — vol-of-vol, tail-risk pricing, bond vol. VRP: implied usually > realised, so option sellers earn a premium.
• Key truth for you: correlations converge toward 1 in stress — "uncorrelated" sleeves sell off together when vol spikes.

• Wyckoff's four phases: accumulation → markup → distribution → markdown — the fractal shape of a trend's life.
• Bubble's five stages (Berman): displacement → boom → euphoria → peak/distress → collapse — each with tell-tale time-series (M2, rate-of-change, sentiment extremes, breadth >80%, insider selling).

• EMH says prices reflect all info — but the joint-hypothesis problem means it can't be cleanly tested, and anomalies (momentum, value, short-term reversal, size, January) persist.
• Adaptive Markets (Lo) — efficiency comes and goes as participants adapt; edges decay and must be refreshed. Fractal Markets (Peters) — different horizons create exploitable structure.

• Prospect theory / loss aversion — losses hurt ~2× more than equal gains, so people let losers run and cut winners early — the mispricing your edges harvest.
• Belief-preservation (confirmation, conservatism, hindsight), information-processing (anchoring, recency, mental accounting), emotional (overconfidence, regret, status-quo) biases.
• Capitulation = collective loss-aversion breaking at once — the fear you buy in a reversal.

• Sentiment (ω) is the gap between price and value; the crowd sets direction in a trend but must be wrong at the extremes.
• Tools: put/call ratio, AAII & Investors-Intelligence surveys (contrarian), COT positioning, short interest, VIX.
• Contrarianism works only at sentiment extremes, never mid-trend.

• P = (F × V)^S — price is fundamentals × valuation, raised to a sentiment power. In a bull, S>1; in a bear, 0<S<1.
• Because price already encodes everything (incl. unknowns), the market is "the best fundamental analyst on the planet" — listening to the trend is hiring that analyst.