The Spinozan Trader

A screen can hold twenty indicators and only one idea. Different colors and decades multiply the appearance of evidence faster than they multiply evidence. The first discipline of this book is not a new oscillator. It is a refusal to let one kind of sentence do the work of another.

A definition fixes how a word is used. Defining an edge as positive expected profit does not demonstrate that you have one. A deduction follows from granted premises: if a position loses half its value, it must double to recover. An empirical claim concerns a dated world. Whether a filter reduced drawdowns after costs is such a claim. The substitution we are here to catch is quiet: “this rule succeeded in these observations” becoming “this rule works.”

You do not need geometry to begin. The Euclidean method here means stating the starting points and showing each step to a conclusion. The word proof does not make an assumption about markets true. “Proposition 3” is an address, not the third step of a system you should run.

X = G − C reads: net profit equals gross profit minus costs. E[X] is the probability-weighted mean of X under an assumed model, not a promise about the next trade. What has been proved is how the quantities relate. What remains unknown is the future distribution of those quantities. The formula cannot supply those inputs.

The book’s first experiment is not a backtest. It is five sentences written before you hunt for confirmation. Name the instrument and exposure. Name the information available before the order. Name the mechanism that could pay you. Name the observation that would reject the mechanism. Name the loss you can survive after costs and adverse execution.

If the third sentence is merely the name of an indicator, the argument has not begun. If the fourth sentence is impossible to write, the idea may be protected from evidence. If the fifth sentence depends on a fill the market has not promised, the account is carrying an assumption you have not acknowledged.

Proposition 1 is almost too plain to skip, which is why it is skipped. Let G be marked profit before modeled costs C. Net profit is X = G − C. Taking expectations where they exist gives E[X] = E[G] − E[C]. What has been shown is an accounting identity. What has not been shown is that your broker’s “commission-free” label makes C equal zero.

Spread, slippage, financing, and shortfall can remain. Equally, an expense already embedded in a fund return must not be deducted a second time. Define the ledger before discussing performance. Taxes and inflation are separate unless the model includes them.

A two-class population has expectancy pW − (1 − p)L − C, where p is the probability of a gross win, W the conditional mean gross gain, L the conditional mean magnitude of a gross loss, and C expected cost. Those letters are inputs. The formula does not estimate them.

Positive expectancy requires p > (L + C) / (W + L) when W + L is positive and the inputs are fixed. Costs raise the break-even winning probability. They cannot be omitted because the strategy wins often. A sample average estimates expectancy. It is not a promise about the next trade.

Proposition 4 is a counterexample, not a market forecast. Strategy A wins 90% of the time, gaining $1 and losing $20, with zero cost. Its expectancy is $0.90 − $2 = −$1.10. Strategy B wins 40% of the time, gaining $3 and losing $1. Its expectancy is $1.20 − $0.60 = $0.60. The higher-win-rate strategy has negative expectancy.

An attractive hit rate may be purchased by occasional large losses. Look at payoff magnitude and tail behavior, not merely the frequency of being right. Classification accuracy in a later chapter fails for the same structural reason.

Capital compounds by multiplication. V_n = V_0 times the product of the wealth factors (1 + r_t). An arithmetic average of returns does not by itself specify terminal wealth. Equal positive and negative percentage returns leave a loss: (1 + a)(1 − a) = 1 − a². A 20% gain and a 20% loss leave 96% of initial capital, in either order.

Recovering a fractional drawdown d requires gain d / (1 − d). A 50% decline requires 100%. A 75% decline requires 300%. These are exact accounting relationships, not forecasts of recovery time. Zero capital is absorbing in a self-financing multiplicative model: once V_t is 0, later factors cannot revive it without an external inflow.

A stop-defined stake limits modeled loss only if execution respects the assumed bound. A planned stop distance is a parameter in an order plan, not a contractual maximum. A sequence of n full-stake losses removes fraction 1 − (1 − f)^n of capital. At 32.5% per bet, five losses remove about 86%. That calculation is not a recommendation to use that stake.

Even a favorable independent strategy can produce a losing streak: q^n is not zero merely because expectancy is positive. The binary log-growth optimum is f* = (pb − q) / b when it lies in the interior. Beyond that point, increasing f lowers expected log growth. Sizing alone does not turn a fair net-return gamble into positive expected logarithmic growth.

Most spectacular backtests begin with a small mistake about time. A daily high is treated as known at the open. A chart marks a Monday pivot with Thursday’s prices. The geometric question is simple: which observations belong to F_t when the order is chosen?

Proposition 22: a signal computed from a completed closing observation must not be credited with an earlier execution. The book’s monthly method computes the signal after the final monthly close and changes exposure at the following session’s open. Existing holdings bear the overnight move. A daily OHLC bar also fails to determine whether the high came before the low. Crediting whichever order is favorable is not a fill model.

Proposition 21: a stop order does not logically guarantee its trigger price as a fill. Let a sell stop trigger at 100. If the next executable price is 95, the fill is 95. That path is compatible with ordinary execution. A guaranteed fill of 100 does not follow.

A stop-limit order changes the permitted execution price but may fail to execute. Neither order type abolishes the tradeoff between price and certainty. Gap, liquidity, venue rules, and instrument structure decide whether the assumed bound is real. Practical ruin often arrives before zero equity: a margin threshold or personal reserve can end the sequence earlier.

Proposition 24: defining a trend does not prove persistence. Define a trend signal as a current observation above an average of past observations. Both a favorable and an adverse next path can share that signal. The definition distinguishes histories. It does not exclude the adverse future. Persistence requires an additional empirical premise.

Proposition 23 is the parent result: a finite price history does not entail a profitable next directional trade. Two admissible paths can share the history and split after the decision. Spinoza’s conatus is a philosophical vocabulary. A measured continuation effect belongs to market research. Do not let the first conceal the missing premise of the second.

Useful conceals several claims. A filter that keeps an anxious investor from trading every hour may improve realized behavior without adding incremental directional information. An expensive feed that predicts the next tick but decays before a fill may contain information and have negative decision value. An indicator is a function I = f(H). An action can be a position, an order, a risk budget, or doing nothing.

Proposition U1: restricting an ideal decision maker to an indicator cannot improve its best attainable expected objective. Deterministic processing does not manufacture an informational advantage over its own complete inputs. A useful compression can still reduce estimation error and inconsistent execution for a finite trader. The universal dismissal of indicators fails. The universal promise of indicators also fails.

Proposition U2: if J = h(I) and h is invertible on the attainable values of I, the decision rules available from I and from J coincide. Changing the scale can make a screen easier to read. It cannot turn the same measurement into independent confirmation.

If position within a channel is x in [0, 1], reporting 100x as a stochastic percentage does not create a new observation. Thresholds 0.8 and 80 describe the same event. Calling ten transformations ten independent sources is not a demonstration. An ablation asks whether removing one component changes a declared outcome after the extra search and trades it introduced.

Proposition U4: classification accuracy need not rank profitable decisions. Strategy A predicts nine small +1 moves correctly, then loses 20. Signed profit is −11 before costs, accuracy 90%. Strategy B predicts four +5 moves and loses 1 on each of six errors. Profit is +14, accuracy 40%. The ranking of accuracy and the ranking of profit reverse.

Prediction can be useful without being directional classification. A variance forecast may inform size. A fill-probability forecast may inform a limit. Four examinations stay distinct: prediction, risk control, execution, and interpretation. Beating the day’s VWAP is not earning investment alpha. A map can be an excellent map of the road already traveled.

A moving average exchanges some sensitivity to the latest observation for a summary of earlier ones. Its lag is not a software defect. Proposition F1: price minus its simple moving average is a weighted sum of recent price changes. The newest change appears most often. “Price above the average” and “weighted recent momentum positive” are two descriptions of the same event under those conventions. They are not two independent witnesses.

On a path rising one dollar per session, a 20-day SMA trails by 9.5 dollars. That gap is explained by the path and the window. It is not a valuation discount and does not imply that price must return to the average. Averaging reduces variance of independent measurement noise. A price is not generally a fixed signal plus independent error. The popular claim that an average “removes noise” needs a model of noise.

A breakout buys a price that recently would have looked expensive. The wager is that a strong movement can contain information about continued adjustment, and that occasional sustained moves can pay for failed entries and costs. That hypothesis needs evidence. The channel only makes the event measurable.

Proposition T1: a close cannot strictly exceed a high channel that includes its own bar, because C_t ≤ H_t ≤ U*_t. A plotted channel may include the present high. A strict close-breakout rule needs a reference that does not swallow the observation it tests. A trailing percentage filter is path-dependent: two paths to the same endpoint can disagree on whether the filter triggered.

An oscillator locates a present price inside a chosen reference range or a scaled memory of changes. RSI, stochastic, and percent-from-average constructions answer “where in this window?” They do not, by definition, answer “what must happen next?” A reading above 70 can be followed by further rise. One steadily rising path is enough to kill “RSI above 70 must be followed by a decline.”

That counterexample does not establish that an RSI strategy always loses. It establishes that the universal reversal claim is false. Usefulness may remain as a description of location, a risk flag, or a component whose incremental value is tested by ablation. Invertible rescalings of the same window are not independent oscillators.

A candle pattern is a specified open-high-low-close geometry. A Fibonacci ratio is a stated fraction of a chosen swing. Neither object, by being named, determines the distribution of the next tradable move. The book’s constructions show that the same pattern geometry can be followed by opposite subsequent returns. The same ratio can be laid on a path that continues and a path that reverses.

A chart is allowed to display a completed geometry at its economic location. A test must timestamp the decision when the geometry is actually known. Retracement tools that use a future swing high to mark an earlier “support” have imported look-ahead unless the decision waits for that high. Elliott labeling that is revised after the fact is interpretation, not a contemporaneous signal.

Momentum is often discussed as one strategy when it describes several comparisons. Time-series momentum compares an instrument with its own past. Cross-sectional momentum ranks instruments against one another. A market in which every asset has fallen can give the two methods opposite exposures.

Proposition M1: the relative winner can have a negative own return. Proposition M2: adding the same constant to every instrument’s formation return leaves ranks unchanged but can flip time-series signs. Intermediate-horizon continuation and very short-horizon reversal can coexist. A monthly winner–loser portfolio is not evidence that buying the last five-minute candle will work. A daily close above an average is not a replication of a volatility-scaled multi-asset futures program.

A backtest is not one experiment when hundreds of rules were tried and one was displayed. The final code can be short while the search was long. Proposition S1: independent repeated testing raises the probability of at least one false rejection. For twenty independent tests at α = 0.05, that probability is about 0.64. Real rules are dependent, so the number is an illustration, not a plug-in correction. Correlation does not make the search disappear.

Out-of-sample evaluation uses observations excluded from the specified fitting process. A repeatedly inspected holdout is no longer fresh. A high Sharpe can mislead: annualizing by √252 does not make autocorrelated days independent. The object of inference must be named. Beating a benchmark, having a positive mean, and meeting a drawdown constraint are different hypotheses.

A profitable idea can be destroyed between the decision and the fill. Execution research asks how to implement a desired position while balancing urgency, price risk, liquidity, and impact. It is a separate source of value from forecasting direction.

Proposition E1: beating a volume benchmark need not make the investment profitable. A trader buys at 99 when VWAP is 100, then values the asset at 90. Execution beat the benchmark by 1; the investment lost 9 before costs. Both statements are true. A simulation that allocates using realized future volume has given itself information unless the live method can respond that way.

Capacity is a claim about how a method’s implementation quality decays as size grows. It is not a compliment. A rule that is executable in a liquid ETF at modest notional can become a different rule in a thin name or a large futures clip. Impact, participation, and the need to split days are part of the method once they change fills.

Regimes are not decorations. A comparison that holds in 1994–2006 can reverse in 2017–July 2026. The book’s monthly filter underperformed buy-and-hold in about 58% of overlapping sixty-month windows in the full sample, and in about 88% after 2007. Overlapping windows share data; they are descriptive, not independent trials. They still prevent a claim of consistent superiority.

The monthly ten-month SPY/cash filter passed its declared historical risk-reduction objective against fully invested SPY. It did not establish that timing was superior to simpler allocations, and it did not earn the highest return. The distinction is the result, not a footnote.

Full sample, January 1994 through July 2026, with the book’s cost and cash conventions: filter CAGR 9.54%, volatility 12.84%, maximum close-to-close drawdown −24.79%. SPY buy-and-hold: 10.83%, 18.76%, −55.19%. Static 50/50: 6.88%, 9.22%, −31.38%. In 2017–July 2026, static 50/50 earned a higher CAGR than the filter with lower volatility and a smaller drawdown. The method held equity on about 78% of daily closes. Holding less equity explains some risk reduction; the result alone does not establish timing skill.

Several kinds of evidence appear in the book. The map exists so a favorable number cannot become a claim about a different strategy. The monthly SPY/cash filter is a retrospective simulation with a specified historical risk tradeoff and substantial benchmark shortfalls. The daily SPY laboratory compares thirteen constructions under common assumptions. It answers how those exact rules differ. It does not baptize the monthly filter.

The stock ORB with daily RVOL ranking is an externally reported 2016–2023 study: a promising published claim requiring independent replication. The long/short NQ ORB is two limited 2026 public-data replays with unresolved contract and execution limitations. Other templates are conditional mathematics and family literature, not new backtested recommendations. Dollar profit, percentage return, and CAGR are not a common ranking scale.

The course was a tribunal. The book is the longer hearing: forty-one chapters, the original propositions, later family demonstrations, a fully specified equity/cash rule, a comparative laboratory, an atlas, worked cases, and an operating protocol. The aim is not a method that cannot lose. Proposition 23 explains why history cannot supply that guarantee. The aim is a method clear enough to execute, limited enough to survive mistakes, and explicit enough to be rejected when it fails its declared purpose.

On a first reading, take method, definitions, expectancy, sizing, and information together, then the calculus of usefulness. Choose a later route by the decision you actually have: specifying a first systematic rule, examining tools on a chart, asking whether a backtest established an edge, or translating ORB research into a test that does not smuggle another market’s return. A good trading book should leave the reader harder to deceive, including by the book itself.