TERMINAL ZERO · SYSTEMATIC DERIVATIVES, EXPLAINED
Six words, in the order the work actually happens. Systematic trading, explained from first principles. This is a teaching site, not a pitch: what a rule actually is, how an order really reaches an exchange, how size is decided, and the vocabulary a systematic trader uses to describe risk. No performance claims, no mystique — just the mechanics.
Illustrative feed · generated client-side · not live market data
Three sentences, each one built only from the one before it. Nothing here requires you to believe anything about markets yet.
A rule is a decision written down so precisely a computer can execute it with no human judgement.
Not “buy when it looks strong” — that is a feeling. A rule names the instrument, the exact condition, the size, and the exit, in terms a machine can evaluate the same way at 09:16 and at 15:20.
A strategy is a set of rules with a defined edge — a reason it should make money more often than not.
The edge has to be a claim about the world: options decay, ranges that break tend to continue, related contracts converge. If you cannot state why it should work, you have a pattern, not an edge.
Quant / systematic trading is running those strategies mechanically, at scale, with statistics deciding when to trade — not emotion or opinion.
Any single trade is close to noise. The point is the thousandth repetition: a small, positive expectancy applied identically, often enough for the distribution — rather than your mood — to show up in the equity curve.
The whole loop, in four moves
Everything else on this page — the Greeks, the sizing, the infrastructure — is detail hung on those four boxes.
Two honest ways to trade the same market. Systematic isn't better — it's different, and the difference is that you can measure it.
A human reads the chart, decides, and clicks.
Rules read the data, decide, and an API clicks.
Depends on the day. The same setup can be taken twice and skipped once.
Identical by construction. The same input produces the same action, always.
Present, and strongest exactly when it costs most — after a loss, or a run of wins.
Removed from execution, but not from the human. It relocates to whether you leave the system running.
Human reaction time: seconds at best, and only one screen at a time.
Milliseconds, on every instrument you subscribe to, simultaneously.
Bounded by attention. Ten positions across three expiries is already a lot.
Bounded by capital, liquidity and margin — not by how many charts you can watch.
Hard to test honestly: the rules live partly in the trader's head and change with hindsight.
Testable by definition, which is also the trap — a rule can be tuned until history flatters it.
A skilled discretionary trader can do things no rule set can: read a situation that has never happened before. A system can do one thing no human can: repeat itself ten thousand times without getting bored, scared, or clever. Pick the one whose failure mode you can live with.
A trade is not a moment. It is a pipeline: five stages that a machine walks in order, thousands of times, without ever wondering whether today feels different.
Prices, quotes and volumes arrive as a continuous stream of ticks. A systematic strategy consumes that stream and nothing else — no headlines, no gut feel. Before a rule can fire, the data has to be clean, timestamped, and shaped identically every single day.
Bad data doesn't produce a bad trade. It produces a confident one.
A signal is a comparison that evaluates to true — a level breaks, an indicator crosses, a spread strays from fair value. It carries no opinion and no hesitation. The same input always produces the same output, which is precisely what makes it testable.
If you can't write the condition down, it isn't a rule yet.
Before anything leaves the machine it asks three questions: how many lots does this size to, does it breach exposure limits, and is there margin to hold it? A failed check isn't a delayed trade — it's no trade. Sizing for survival happens here, not afterwards.
Position sizing is where most of the outcome is actually decided.
The order goes to the broker's API and rests or crosses on the exchange's order book. What comes back is a fill — sometimes partial, often a shade worse than the price that triggered it. That gap is slippage, and it is a real, recurring cost of the edge.
The price you signalled at is not the price you own.
P&L is marked continuously, stop and target logic runs on every tick, and positions are reconciled against what the broker actually holds. A kill switch stands ready to flatten everything if the market — or the machine — misbehaves.
Exits and reconciliation are the unglamorous half that keeps you solvent.
01 · Data
Prices, quotes and volumes arrive as a continuous stream of ticks. A systematic strategy consumes that stream and nothing else — no headlines, no gut feel. Before a rule can fire, the data has to be clean, timestamped, and shaped identically every single day.
Bad data doesn't produce a bad trade. It produces a confident one.
02 · Signal
A signal is a comparison that evaluates to true — a level breaks, an indicator crosses, a spread strays from fair value. It carries no opinion and no hesitation. The same input always produces the same output, which is precisely what makes it testable.
If you can't write the condition down, it isn't a rule yet.
03 · Risk check
Before anything leaves the machine it asks three questions: how many lots does this size to, does it breach exposure limits, and is there margin to hold it? A failed check isn't a delayed trade — it's no trade. Sizing for survival happens here, not afterwards.
Position sizing is where most of the outcome is actually decided.
04 · Execution
The order goes to the broker's API and rests or crosses on the exchange's order book. What comes back is a fill — sometimes partial, often a shade worse than the price that triggered it. That gap is slippage, and it is a real, recurring cost of the edge.
The price you signalled at is not the price you own.
05 · Monitor & exit
P&L is marked continuously, stop and target logic runs on every tick, and positions are reconciled against what the broker actually holds. A kill switch stands ready to flatten everything if the market — or the machine — misbehaves.
Exits and reconciliation are the unglamorous half that keeps you solvent.
Six families, each with a stated edge and a stated way of failing. If a strategy can't tell you how it loses, you don't understand it yet.
// the common thread
None of these are predictions.
Each one is a claim about structure — decay, convergence, continuation — plus a rule that harvests it and a limit that survives being wrong. The skill is in the limit, not the claim.
swipe →
An option's price is a function of a handful of inputs. The Greeks are simply the slopes of that function — how much the price moves when one input moves. Drag something and watch.
Illustrative Black–Scholes on a 100-point index. Values are per unit, not per lot.
An option price is not a prediction. It is the cost of transferring a defined risk to somebody else, assembled from five inputs — four of which you can look up.
The whole identity
price = intrinsic + extrinsic
Intrinsic is what the option would be worth if it expired right now — the in-the-money part, and nothing else. It is pure arithmetic on the spot and the strike.
Extrinsic is everything above that: the price of the time left and the movement expected within it. It is the entire argument. Theta erodes it, vega re-prices it, and at expiry it is exactly zero. A seller is selling extrinsic value; a buyer is renting it.
Put–call parity
C − P = S − K·e−rT
A call and a put on the same strike and expiry are locked together by arbitrage. Hold a call and sell a put and you have synthesised the underlying itself — so if the relationship breaks, the trade is free money and it gets closed in seconds.
This matters more than it looks. It means calls and puts are not separate opinions about direction; they are the same instrument rearranged. Any story that requires a call to be cheap and a put to be expensive at the same strike is not a story about volatility — it is a mispricing that does not exist.
// four you can look up, one you cannot
S
Spot
where the underlying is now
K
Strike
the price you have the right to trade at
T
Time to expiry
how long the optionality lasts
σ
Volatility
how much it is expected to move — the only unobservable
r
Rate
the cost of carrying the position
Spot, strike, time and rates are facts. Volatility is a forecast, and it is the only input anybody argues about — which is why option traders describe themselves as trading volatility rather than direction. Everything else in the price is bookkeeping.
// the model runs backwards
Nobody prices options by feeding a volatility number into Black–Scholes and accepting the answer. They observe the market price and solve backwards for the volatility that would justify it. Implied vol is the output of that inversion — a way of restating a price in units that can be compared across strikes and expiries.
Black–Scholes assumes volatility is a single constant. If that were true, every strike would imply the same number and the curve opposite would be flat. It never is. Out-of-the-money puts consistently imply more volatility than calls, because crashes are faster and deeper than rallies and everyone knows it. The shape is the market pricing in the fat tail the model leaves out.
Implied volatility by strike · illustrative equity index
Downside strikes cost more per unit of volatility. That is the skew, and it is the market's standing bid for crash protection.
// why any of this is hard
Eugene Fama's hypothesis comes in three strengths, each a claim about what is already in the price. Treat it not as a law but as the null hypothesis: the thing your strategy has to disprove before you are allowed to believe in it.
Weak form
Prices already reflect all past prices and volume.
Semi-strong form
Prices reflect all publicly available information.
Strong form
Prices reflect all information, public and private.
The joint hypothesis problem
You can never test market efficiency on its own. Any test asks whether returns beat a model of what returns should be, so a rejection means either the market is inefficient or your model of fair return is wrong — and the data cannot tell you which. Fama pointed this out himself.
The Grossman–Stiglitz paradox
If prices already reflected everything, nobody would be paid to do research — and if nobody did research, prices would stop reflecting anything. Markets must therefore stay slightly inefficient, by exactly enough to pay the people making them efficient. That margin is where every edge on this site lives.
Every field builds a private vocabulary, and trading's is worse than most. Here is the whole working set in one sentence each — search it, or browse by family.
61 terms
Strategy code, a signal engine, a risk layer, a broker API — and a loop that brings reality back so the machine knows what it actually holds.
No manual clicks anywhere in that loop. The human writes and supervises the rules; the machine trades them.
Systematic trading is not a novel idea. Two of the oldest, largest industries on earth run on exactly the same three parts: a small structural edge, an enormous number of repetitions, and a hard cap on what any single outcome can cost.
// the wheel does not need to know
A single-zero wheel has 37 pockets. A straight-up number pays 35 to 1, but the true odds against it are 36 to 1. That one-pocket gap — the green zero — is the entire business. It works out to 2.70% of every rupee staked, and it applies whether the ball lands on your number or not.
The house loses individual spins all night long. It is indifferent, because it is not playing your session. It is playing every session, forever.
Realised house take vs number of spins
Five simulated runs of four thousand spins. Early on the house is sometimes losing badly — that is variance, and it is why a short session tells a gambler nothing. Push the count high enough and every run is pinned to the same 2.70%. The casino is not lucky. It is patient, and it has a bigger sample than you do.
A positive count means the shoe is rich in tens and aces. The edge has moved to the player — which is the moment the casino changes the rules.
// the exception that proves the rule
Cards leave the shoe and do not come back, so the deck has a memory. Count the low cards going out and you know when what remains is rich in tens and aces — and in those moments the advantage genuinely belongs to the player.
The casino's response is instructive: more decks, earlier shuffles, and a polite escort to the door. An edge that becomes widely known gets competed away or shut down. Markets do the same thing, only faster and without the bouncer.
// the same machine, wearing a suit
A premium is expected loss, plus expenses, plus margin. Nobody at the underwriting desk believes they know whether you will crash your car. They know what a hundred thousand drivers like you do in aggregate, and they price the distribution. The measure of whether it worked is the combined ratio: claims plus expenses over premiums. Under 100% and the book made money on the risk itself.
The parallels run all the way down. Policy limits and deductibles are position sizing. Reserves are the capital buffer that keeps a bad year from being the last one. Reinsurance is a tail hedge — paying someone else to take the extreme. And a hurricane is the thing that breaks the whole model, because it turns a hundred thousand independent policies into one enormous correlated bet.
None of these businesses forecasts the next outcome. The advantage is built into the rules of the game before anyone plays.
CasinoA European wheel has 37 pockets. A straight-up number pays 35 to 1, but the true odds against it are 36 to 1. The house keeps that gap — 2.70% of everything staked — whatever the ball does.
A 2.70% edge is invisible in ten spins and inescapable in ten million. The law of large numbers is the entire business model.
CasinoThe house loses hands constantly and does not care. It needs volume, not wins — enough spins for realised outcomes to converge on the maths.
Edge only pays out if you are still solvent to collect it. Every one of these businesses puts a ceiling on single-event damage.
CasinoTable maximums are position sizing. A bet cap limits how much variance one player can impose on one table in one night.
Diversification is the machine's load-bearing assumption. Everything breaks on the day every outcome becomes the same outcome.
CasinoA table full of players betting the same number turns thousands of independent spins into one wager. Correlated action is the one thing a pit boss genuinely fears.
The uncomfortable half of this: the arithmetic that makes the house rich makes the player poor, by exactly the same mechanism.
A negative edge repeated ten thousand times is not bad luck — it is a calculation running to completion. Nothing about sitting on the other side of the table changes the maths; it only changes the sign. Which is the whole reason to be able to state what your edge is, why it should exist, and whether it is positive — before you put size on it. If you cannot, you are not the house. You are the sample the house is collecting.
A positive edge is necessary and nowhere near sufficient. What decides whether you collect it is how much you stake, and whether you are still solvent after the worst run the maths guarantees you will eventually see.
// the streak is not a malfunction
Flip a fair coin a thousand times and somewhere in there is a run of about ten tails. Nothing is broken when it happens; a thousand-flip sequence without such a run would be the anomaly. The expected longest losing run is ln(N) / ln(1/q) — where N is the number of trades and q the probability of a loss.
This is the single most useful number a systematic trader can hold in their head, because it converts “my system stopped working” into a testable claim. If your backtest says 45% win rate and you are eleven losses deep, you are exactly where you should be. Abandoning a strategy inside its expected run is how a positive edge gets thrown away.
Expected longest losing run · 1,000 trades
| Win rate | Longest run |
|---|---|
| 60% | 7.5 losses |
| 55% | 8.7 losses |
| 50% | 10.0 losses |
| 45% | 11.6 losses |
| 40% | 13.5 losses |
| 35% | 16.0 losses |
| 30% | 19.4 losses |
// the equity curve you got was one draw
Your historical equity curve is one path the dice happened to take. Keep the same edge and reshuffle the order of the trades and you get a different curve — sometimes a much uglier one. Simulating thousands of those orderings turns a single number into a distribution, and the number you care about is not the average outcome but the bad tail: how deep does the drawdown get in the unlucky 5%, and could you sit through it?
24 orderings of the same 120 trades · 52% win rate · 1:1 · 2% risked
Kelly fraction · f* = (bp − q) / b
| Odds | Win | Full | Half |
|---|---|---|---|
| 1 : 1 | 52% | 4% | 2% |
| 1 : 1 | 55% | 10% | 5% |
| 1 : 1 | 60% | 20% | 10% |
| 2 : 1 | 40% | 10% | 5% |
| 3 : 1 | 35% | 13.3% | 6.7% |
// the growth-optimal bet
Kelly maximises the long-run growth rate of capital, and it is provably optimal in that specific sense — no other staking rule beats it asymptotically. It is also the only rule on this page that requires a positive edge as an input, which is exactly what makes it useful: you cannot apply it without first stating what your edge is.
The catch is variance. Full Kelly is violent — drawdowns of 50% are routine, and it assumes your estimate of the edge is exact, which it never is. Overestimate the edge and you are over-betting immediately. Because growth is flat near the optimum, half Kelly gives up only about a quarter of the growth for roughly half the volatility, which is why most people who use it at all use a half or a quarter.
// having an edge is not the same as surviving
Betting one unit at a time with a genuine edge, the probability of losing everything is (q/p)^N, where N is how many units of bankroll you hold. The edge barely moves that number. The unit count moves it enormously.
At a 52% win rate with ten units you are close to a coin flip on total ruin — 45%. The same edge with fifty units puts it under 2%. Nothing about the strategy changed; only the size did. This is the entire argument for position sizing, and it is why the risk gate sits before the order goes out rather than after.
Probability of ruin · (q/p)^N
| Win | Units | Ruin |
|---|---|---|
| 52% | 10 | 44.91% |
| 52% | 20 | 20.17% |
| 52% | 50 | 1.83% |
| 55% | 10 | 13.44% |
| 55% | 20 | 1.81% |
| 55% | 40 | <0.1% |
| 60% | 10 | 1.73% |
| 60% | 20 | <0.1% |
// the asymmetry nobody plans for
Losses and gains are not symmetric, because the gain is computed on a smaller base. Down 50% you need to double just to get level. This one table is the reason experienced traders talk about risk before they talk about return — and the reason a strategy with a smaller edge and a shallower drawdown often compounds better than a flashier one.
Gain required to recover a drawdown · gain = loss / (1 − loss)
Bars are square-root scaled — at true scale the 90% row would be seventeen times the 20% row and the table would be unreadable, which is itself the point.
// what the players at the table are doing
Traders reinvent all of these, usually without knowing the names. They are worth knowing precisely because they are so tempting — and because the honest verdict on almost all of them is the same. A staking system reshapes the distribution of outcomes. It cannot touch the mean. No sequence of bet sizes turns a negative edge positive; it only decides whether you lose in many small pieces or one large one.
The default everywhere
18th-century France
The mirror of Martingale
Named for the French mathematician
The sequence, applied to stakes
19th-century England
Documented by Allan Wilson, 1965
Standard in money management
John L. Kelly Jr., Bell Labs, 1956
Only two of the nine survive contact with arithmetic, and both of them start by asking a question none of the others do: what is your edge, and how sure are you?
Most of a quant's working life is not spent finding strategies. It is spent trying to prove that the one in front of them is not an accident — and these are the tools that do the disproving.
Before any model, know what the distribution actually looks like. Market returns are not normal, and every technique that assumes they are will understate your risk.
Is the distribution lopsided?
=γ₁ = E[(X − μ)³] ⁄ σ³
Third moment. Negative skew means many small gains and rare large losses — the signature of short-option and carry strategies.
The trap
A strategy can post a beautiful Sharpe for years purely by hiding risk in negative skew.
How often do extreme moves happen?
=γ₂ = E[(X − μ)⁴] ⁄ σ⁴ − 3
Fourth moment. Excess kurtosis above 0 means the tails are heavier than a normal distribution predicts.
The trap
Financial returns routinely show excess kurtosis of 5–10. A 'six-sigma' day is not once-a-millennium; it is roughly annual.
Are returns normally distributed?
=JB = n⁄6 · (S² + (K − 3)² ⁄ 4)
Tests skew and kurtosis jointly against the normal. The QQ plot shows you the same thing by eye.
The trap
They almost always reject normality. The useful question is not whether but how badly, and where.
None of these tools will find you a strategy. Every one of them exists to take one away — and a process that never kills anything is not a research process, it is a way of laundering hope into conviction.
Anybody can produce a rising equity curve from ten years of history — the data will happily confess to anything if you torture it long enough. The whole discipline is in the tests you run afterwards to find out whether the curve means anything, and in being willing to throw the system away when it does not.
The pipeline
Skipping one does not save time. It moves the discovery of the problem to the point where it costs money.
One sentence, before any data: who is on the other side of this trade, and why do they keep taking it? An edge is somebody else's cost, constraint or hurry.
What goes wrong
If you cannot name the loser, you are not testing a hypothesis — you are searching for a coincidence.
Adjusted for splits and dividends, corrected for expiry rolls, free of survivorship bias, timestamped in the exchange's clock, with the bad prints and the gaps marked rather than silently filled.
What goes wrong
A bad tick that never traded becomes a fill in a backtest. Most spectacular backtests are made of prices nobody could have got.
Entry, exit, stop, filter, position size and what happens when two signals collide — all of it written so that the same input can only ever produce one action.
What goes wrong
Any rule that still needs you to look at the chart is a discretionary system wearing a backtest as a costume.
Brokerage, exchange fees, STT and stamp duty, GST, the bid-ask spread you actually cross, slippage that scales with your size, and financing on anything you carry.
What goes wrong
Costs are the single most common reason a live system underperforms its backtest. Model them pessimistically or they will be modelled for you.
The engine walks forward one bar at a time and may only use information that existed at that instant — including the fact that a bar's close is not known while the bar is forming.
What goes wrong
Look-ahead bias. Using a daily close to decide a trade you booked at that same close is the classic version, and it is worth several imaginary percent a year.
Plot cumulative P&L and, underneath it, the underwater curve — how far below the previous high-water mark you are, and for how long.
What goes wrong
Everyone looks at the line going up. The number that ends careers is how long the line went sideways.
A handful of metrics you can compare across systems: return, drawdown, risk-adjusted return, expectancy per trade, and how much of it survives costs.
What goes wrong
One metric alone is always gameable. Sharpe hides tail risk, win rate hides payoff, CAGR hides the path.
Hold back a slice of history, build and tune on the rest, and look at the held-back slice exactly once.
What goes wrong
Look at it twice and it is in-sample. Every peek spends part of the only unbiased evidence you have.
Nudge every parameter. Shuffle the trades. Delete the best month. Split by regime — trending, choppy, high-volatility, low-volatility, pre- and post-2020.
What goes wrong
A system that only works with one exact parameter set, or only in one regime, has not been discovered. It has been drawn around the data.
Run the finished rules on data as it arrives, without money, logging every signal, fill assumption and latency. Weeks to months, depending on how many trades it makes.
What goes wrong
This is the first test the data cannot have been fitted to, because the data did not exist when you built the rules.
Real money, smallest tradeable size, full monitoring, a kill switch, and a written rule for what would make you turn it off.
What goes wrong
The purpose of the first live months is not profit. It is to find the gap between your fill assumptions and your actual fills.
Reading the result
Every metric on this list is a summary, and every summary throws information away. Read them as a set — the interesting information is usually in the disagreement between two of them.
| Metric | What it says | Where a real system lands | How it lies |
|---|---|---|---|
| CAGR(Vₑ ⁄ V₀)^(1⁄yrs) − 1 | The compound rate the equity curve actually grew at. | Anything you would be happy to earn for a decade. | Says nothing about the path. Two systems with the same CAGR can be unrecognisably different to live through. |
| Max drawdownmin(Vₜ ⁄ max V₍≤ₜ₎ − 1) | The worst peak-to-trough fall on the whole curve. | Assume the live one is 1.5× the backtested one. | It is a single historical observation, not a limit. The worst drawdown in your future is almost always larger than the worst in your past. |
| Calmar / MARCAGR ⁄ |max DD| | Return per unit of the worst pain. The most honest single number on this list. | Above 0.5 is workable; above 1 is good; above 3 usually means a bug. | Both inputs come from single extreme points, so it is noisy on short histories. |
| Sharpe ratio(E[r] − rբ) ⁄ σ(r) | Excess return per unit of volatility. | 1 is respectable after costs; 2 is excellent; above 3 in retail-accessible markets deserves suspicion. | Punishes upside volatility as much as downside, and assumes returns are roughly normal — which strategies that sell options are specifically designed not to be. |
| Sortino ratio(E[r] − rբ) ⁄ σ(downside) | The same idea, but only downside deviation counts. | Read alongside Sharpe, never instead of it. | A large Sortino-to-Sharpe gap tells you the returns are skewed; it does not tell you which way the skew will resolve. |
| Expectancyp·avg win − (1 − p)·avg loss | The average rupee outcome of one trade, after costs. | Must exceed your round-trip cost by a comfortable multiple. | An expectancy driven by a handful of enormous winners is not repeatable. Check the median trade too. |
| Profit factorΣ wins ⁄ Σ |losses| | Rupees won per rupee lost. | 1.3–1.8 after costs is a real, tradeable system. | Above 2.5 on a few hundred trades usually means the sample is too small or the costs are too kind. |
| Win ratewins ⁄ trades | How often you are right. | Meaningless in isolation. 35% and 85% are both fine. | The most seductive number in trading. Any system can push it to 90% by cutting winners early and letting losers run. |
| Payoff ratioavg win ⁄ avg loss | The other half of win rate. The pair is what matters. | Win rate × payoff must clear 1 by enough to pay costs. | Averages hide the one loss that was eight times the others. |
| Trade countn | How much evidence you actually have. | Under 100 trades, you know almost nothing. Several hundred is a starting point. | Ten years of daily data sounds like a lot. If the system trades monthly, that is 120 observations. |
| Longest flat periodmax gap between equity highs | The longest stretch you would have gone without a new high. | However long it is, plan to sit through longer. | Nothing lies here. This is the number that decides whether you can actually keep running the system. |
| Cost ratioΣ costs ⁄ Σ gross P&L | How much of the gross edge the market takes back. | Under a third is healthy. Over a half is fragile. | It is the metric that degrades fastest as you scale, because slippage grows with size while the edge does not. |
Benchmarking
These two curves finish at almost the same place. One is a business; the other is a story about a business.
A result worth deploying
A result that will not survive
Benchmark against
If the index returned more with less drawdown, your system is an expensive way to underperform a SIP.
Benchmark against
Replace your entry signal with a coin flip and keep everything else. If results barely change, your edge was in the exit or the sizing, not the signal you believe in.
Benchmark against
Destroy the time structure and re-run. Any performance left is what the rules extract from noise — and that is your false-positive baseline.
Benchmark against
The only honest comparison is in-sample versus out-of-sample on the same rules. A fall of more than about a third is a warning; a fall to zero is an answer.
The system itself
Not the return. Two systems can post identical numbers and only one of them will still be working in three years.
Fragile
Durable
Nine parameters, each tuned to a specific value
Two or three parameters, each working across a wide neighbourhood
The rationale was written after the backtest
The rationale was written before, and the test could have refuted it
Profit concentrated in a handful of trades
Profit spread across hundreds of trades and several years
Works in one regime and quietly loses in the others
Positive or flat in every regime, best in the one it was designed for
Backtest and live results diverge and nobody knows why
Live fills are reconciled against simulated fills every day
Costs modelled as a flat, optimistic number
Costs modelled per-instrument, with slippage scaling in size
One equity curve, and it is beautiful
A distribution of equity curves, and the bad tail is survivable
No stated condition for turning it off
A written kill rule agreed before the first rupee is risked
Out-of-sample & forward testing
The single most useful thing you can do to a backtest is refuse to look at part of the data until you have finished building.
Naive
Hold-out
Walk-forward
The slice you build and tune on. Everything you learn here is contaminated by the fact that you learned it here.
A slice held back and untouched. It answers one question — does this generalise? — and it can only answer it once, because every look leaks information into the next revision.
Fit on a window, trade the next window, roll forward, repeat. It produces a continuous out-of-sample track record and it mirrors how you would actually re-fit a live system, which is why it is the standard.
Paper trading on data that did not exist when the rules were written. It is the only test that cannot be contaminated, and the only one that also tests your infrastructure.
Optimisation & curve fitting
Searching a parameter space is not cheating; you have to pick numbers somehow. It becomes curve fitting the moment the number you picked is the only one that works.
With enough parameters you can fit any history perfectly, and the fit will contain exactly zero information about the future. The tell is always the same: performance falls off a cliff one step away from the chosen settings.
A useful rule: if moving a parameter by 20% in either direction destroys the result, you have not found a setting — you have found a fluke, and the setting is a memory of where it happened.
Monte Carlo
History dealt your trades in one particular order. Nothing about the future guarantees a similar order — so shuffle them, thousands of times, and look at the range of businesses you might have been running.
Twenty-two reorderings of one set of trades. The backtest you saw was a single line through this cone — and there is no reason it was the middle one.
Shuffle the sequence of realised trades and rebuild the curve. Same trades, same expectancy, wildly different drawdowns.
Bootstrap the trades, allowing repeats and omissions. This asks what happens if the future draws from the same distribution but not the same sample.
Shift each entry by a bar or two. A system that only works on the exact bar is an execution fantasy.
The number that matters is the bad tail: the drawdown you would have suffered in the unlucky 1-in-20 ordering. Size the position from that, and the median takes care of itself.
Finalising
Not a wish list. Any single line unticked is a reason to keep the system in paper.
The hypothesis was written before the test, and it could have been refuted.
Costs are modelled pessimistically and the system still clears them by a wide margin.
Out-of-sample results were looked at once, and they held.
Every parameter sits on a plateau, not a peak.
The Monte Carlo bad tail is a drawdown you could actually sit through at your intended size.
The forward test produced live signals that match the simulated ones.
Position size is set from the Monte Carlo tail, not from the backtested drawdown.
There is a written rule for when this system gets switched off.
After the analysis
The backtest is the optimistic case by construction, because it was built by someone who already knew what happened. Here is the gap you should budget for.
What the backtest implies
What live data does
Fills at the price on the chart
You cross the spread on entry and again on exit, and in illiquid strikes the spread is the whole edge.
Every signal gets executed
Rate limits, rejected orders, margin shortfalls, frozen quantity limits and the odd exchange halt mean some do not.
Performance in line with the backtest
A live system that delivers 60–70% of its backtested risk-adjusted return is doing well. Plan the position size around that number, not the backtest's.
Drawdown similar to the tested worst case
Expect to exceed the backtested max drawdown eventually. It is a sample maximum, not a bound.
The edge persists
Edges decay. Crowding, exchange rule changes, tick-size changes and regime shifts all end strategies. Track rolling expectancy, not lifetime expectancy.
Your job is done at deployment
Roughly half the work of running systems is reconciliation, monitoring and knowing when to stop — which is the part nobody backtests.
The point of a backtest is not to find out how much money a system would have made. It is to find out how it fails — and if you cannot make it fail on paper, you have not tested it, you have admired it.
Not from a secret indicator. From doing one statistically sound thing thousands of times without flinching — and from surviving the stretches where it doesn't work.
Illustrative — a shape of work, not a track record
A backtest is a hypothesis about the past. Keep data out of sample, walk it forward, and assume the live result will be worse than the test — because it always includes costs the test forgot.
Every positive-expectancy system spends most of its life below its own high-water mark. The drawdown isn't the system breaking; it's the system working, in the part nobody screenshots.
Expectancy only pays out if you're still there to collect it. Sizing decides whether a normal losing streak is an inconvenience or the end of the account.
Almost all the work is research, costs, sizing and risk. The entry rule is usually a single line of code — and rarely the reason a system succeeds or fails.
Built by a non-discretionary quant trader working Indian equity derivatives entirely through automated API infrastructure.
Most trading education sells an outcome. This one tries to hand over a vocabulary and a mental model instead: what a rule is, where risk is actually decided, and what happens in the milliseconds between a condition turning true and an order resting on an exchange.
Nothing here is advice, a signal service, or a claim about returns. Every number and chart on this page is generated in your browser to illustrate a shape — decay, convergence, drawdown — not to report a result. Derivatives can lose more than you put in; understand the mechanics before you risk anything on them.
Built in public, one rule at a time.
// keep learning