01 / THE 15-YEAR CHART
A comfort blanket, not an edge.
Every trader inherits the same assumption: more history means more truth. Pull up 15 years of daily candles, backtest the system across two crashes and three bull runs, and if it survives, it must be robust.
At AtomicTrades we think this gets the problem backwards. A 15-year backtest proves your idea worked in fifteen different markets, most of which no longer exist. The market you are trading tomorrow is the one from the last six months, and in some cases the last six weeks.
This post makes the case with five arguments, drawn from regime-switching econometrics, volatility forecasting, market microstructure, alpha-decay research, and the mathematics of card counting. We also give the honest counterargument at the end, because long history is not useless. It is just useful for a different job than most traders give it.
02 / REASON 1 · THE MARKET IS NOT STATIONARY
Old data describes a market that no longer exists.
Every long backtest rests on one silent assumption: the statistical rules that generated prices in 2011 are the same rules generating prices today. That assumption has a name, stationarity, and financial data fails it almost everywhere you look. Distributions drift, parameters wander, and structural breaks arrive without warning.
A 2025 paper from Columbia (Capponi, Huang, Sidaoui, Wang and Zou, The Nonstationarity-Complexity Tradeoff in Return Prediction) put a number on this. Testing models on 17 industry portfolios from 1987 to 2016 with training windows from 1 month to 256 months, they found that during regime changes (the Gulf War recession, the dot-com bust, the 2008 crisis) simple models trained on short windows beat complex models trained on long or expanding windows.
14–23%Better out-of-sample R² when the window length adapts
+31%Higher cumulative trading returns than the best fixed window
1–256 MOTraining windows tested, 17 industry portfolios, 1987 to 2016
The mechanism is simple. Prediction error has three parts: how wrong your model’s shape is, how noisy your estimate is, and how much the world has changed since your oldest data point. Adding years of history shrinks the second term and inflates the third. Past a certain point, every extra year of data makes you more confident about a market that has already left.
PREDICTION ERROR AGAINST WINDOW LENGTHBEST WINDOW ≈ 14 MONTHS
Noise (lavender) shrinks with more data. Drift (pink) grows with it, because the oldest bars come from a market that has moved on. The total (cyan) has a floor. Slide the market’s speed of change: the faster it moves, the shorter the best window. Only a market that never changes rewards using all of its history.
For a day trader this matters even more than for an academic. You are not estimating a 20-year risk premium. You are estimating how this instrument behaves this week. A 15-year lookback drowns that signal in fourteen years of noise from other regimes.
03 / REASON 2 · THE MARKOV PROPERTY
Tomorrow depends on today, not on 2011.
A Markov process has one defining feature: the next state depends only on the current state, not on the full path that led there. If you know where the market is now and which regime it is in, the rest of the history adds nothing to your forecast.
This is not a fringe idea. James Hamilton’s 1989 regime-switching model, the foundation of most institutional regime work, assumes exactly this. The market sits in one of a few hidden states (calm, stressed, trending, mean-reverting), each with its own mean and volatility, and it jumps between them according to a transition matrix. The model does not care what happened ten years ago. It cares which state you are in today and how sticky that state is.
A YEAR FROM A THREE-STATE MARKOV MARKETSTATE · CALM
CALMregime right now
0bars in this regime
33bars a regime lasts, on average
Three hidden states, each with its own drift and volatility, drawn in cyan (calm), violet (trending) and pink (stressed). The stickiness is the diagonal of the transition matrix: how likely the market is to stay where it is from one bar to the next. Nothing in the simulation looks further back than the current bar.
Hedge funds build on this directly. Markov-switching GARCH models have been the standard tool for analysing hedge fund return regimes since the mid-1990s, because they capture the things long-run averages hide: fat tails, volatility clustering, and the fact that correlations change when the regime changes. Regime detection in practice means estimating the current state from recent data, then adjusting position size and strategy parameters to that state, not to the 15-year average.
If the process is close to Markov, the information content of a price bar decays fast with age. The last six months tell you which regime you are in. The six years before that are mostly a record of regimes you are not in.
04 / REASON 3 · VOLATILITY CLUSTERS
The industry’s own risk models already ignore old data.
The most robust empirical fact in finance is volatility clustering: big moves follow big moves, quiet days follow quiet days. Today’s volatility is best predicted by yesterday’s, not by the average of the last 3,000 sessions.
The risk models banks actually run encode this with exponential decay. RiskMetrics uses a decay factor of 0.94 for daily data. Run the arithmetic and the implication is stark:
WEIGHT OF A PRICE BAR BY AGEHALF-LIFE · 11 DAYS
| Age of the price bar | Weight the model gives it | |
|---|
RISKMETRICS USES 0.94 FOR DAILY DATA
The bar for each row is on a logarithmic scale: every notch to the left is ten times less weight. Even at a slow 0.99 decay, a bar from a year ago carries under a tenth of a percent.
The half-life of information in the standard daily volatility model is about 11 trading days. Anything older than six months contributes effectively nothing. The 15-year chart on your screen is, by the industry’s own risk arithmetic, 99.9 percent irrelevant to tomorrow’s expected range.
GARCH models, the more general version, are re-estimated daily on rolling windows of recent data at NYU’s V-Lab for the same reason: parameters fitted on old data stop reflecting current dynamics. Nobody who forecasts volatility for a living uses an equal-weighted 15-year window. They use the last few weeks, weighted toward the last few days.
05 / REASON 4 · THE PARTICIPANTS CHANGED
The crowd changed, so the game changed.
A chart from 2012 records the behaviour of a crowd that no longer trades. Over the last few years the composition of order flow in US markets has shifted more than in the previous twenty, and a long backtest averages across all of it as if it were one population.
The SEC’s own April 2026 staff report on options market structure lays out the scale: the number of unique underliers with traded options rose 144 percent from 2012 to 2025, OPRA quote messages went from roughly 9 billion a day in 2017 to 131 billion a day by December 2025, and trading on the day of expiration (0DTE) grew from 19.6 percent of all options volume at the start of 2022 to 28 percent by 2025. On expiration days, individual customers account for 73.5 percent of non-market-maker volume, versus 54.3 percent on other days. The retail crowd, armed with same-day options, is now the marginal participant on a large share of trading days.
HOW MUCH THE ORDER FLOW MOVEDSEC STAFF REPORT · APRIL 2026
Three measures of who is trading and how. Each pair is drawn to its own scale; the labels carry the numbers.
In the S&P 500 complex specifically, the shift is more extreme. By September 2025, 0DTE contracts exceeded 60 percent of US stock options volume, up from around 15 percent a few years earlier. That changes the mechanics of the underlying: dealers hedging short-dated gamma produce concentrated bursts of buying and selling, intraday mean reversion speeds up until it snaps, and a falling VIX can coexist with rising intraday fragility.
WHAT A 15-YEAR BACKTEST IS MADE OF12 OF 15 YEARS · A CROWD THAT IS GONE
DIM · A MARKET WHERE 0DTE BARELY EXISTEDLIT · THE CROWD YOU TRADE AGAINST
Think about what this does to a 15-year backtest of an intraday strategy. Roughly 80 percent of the sample comes from a market where 0DTE barely existed, retail options flow was marginal, and dealer gamma hedging was spread across quarterly and monthly expiries. The pattern your system found in 2014 was produced by participants who have been replaced. The last 6 to 12 months are the only slice of data generated by the crowd you are actually trading against.
06 / REASON 5 · THE EDGE WAS ALREADY DYING
Long histories average away the edge.
Every edge decays. The moment a pattern becomes known, capital crowds into it and the return gets competed down. A 15-year backtest does not show you the edge as it exists today. It shows you a blend of the edge when it was fresh, the edge as it was discovered, and the edge after it was arbitraged away, all averaged into one misleading number.
MOMENTUM · ROLLING SHARPE RATIO, STYLISED1.5 IN THE MID-1990s · 0.25 TODAY
Drawn from the figures in the text, not from a data feed: the shape, not the wiggles, is the point. The dashed line is a 15-year average of the curve: it describes no year that actually happened.
The evidence here is unusually clean:
−58%Return of 97 published predictors after the paper came out (McLean and Pontiff, 2016)
−2.1 to −3.1 PTSLive performance a year below the backtest, 1,726 marketed strategies (2026)
−3.5 to −4.5 PTSFor strategies launched after a hot run for their factor; cold launches improved
- McLean and Pontiff (2016) tracked 97 published return predictors and found their returns fell by roughly 58 percent after the academic paper came out. The momentum factor returned about 10 percent a year in the 1990s and is closer to 2 percent now. Its rolling Sharpe ratio fell from around 1.5 in the mid-1990s to around 0.25 today, and the decay accelerated after 2015 as factor ETFs made the trade available to everyone.
- A 2026 study of 1,726 commercially marketed trading strategies from ten global institutions found that live performance dropped 2.1 to 3.1 percentage points a year below the backtest, and that once you benchmark against peers, the backtest explains almost nothing. The author’s conclusion: marketed backtests mostly capture the favourable regime present before launch, not skill. Strategies launched after a hot run for their factor decayed by 3.5 to 4.5 points. Strategies launched into a cold regime actually improved slightly.
Read that second finding again, because it is the whole argument in one line. The long backtest is not measuring your strategy. It is measuring the regime. And the regime is exactly what a short lookback captures and a long one blurs.
There is a second, quieter problem with long windows: they invite overfitting. Fifteen years of daily data give you thousands of bars to tune parameters against, and a strategy that survives all of them has usually been bent to fit them. A six-month window leaves fewer knobs to turn and forces you to ask the only question that matters: does this work in the market I have right now?
07 / THE GAMBLER’S VIEW
Count the shoe in front of you.
At AtomicTrades we say day trading is gambling, and we mean it as a description, not an insult. So it is worth asking how professional gamblers think about history.
A card counter at a blackjack table does not care what came out of the shoe last Tuesday. The edge comes from one thing: the composition of the cards remaining in the shoe right now. The counter tracks a running count of high and low cards dealt, converts it to a true count by dividing by the decks remaining, and bets big only when the current shoe is rich in tens and aces. Each true-count point is worth roughly half a percent of edge. The count resets to zero every time the shoe is shuffled. Ten years of past shoes contribute nothing.
SIX-DECK SHOE · HI-LO COUNTBET SMALL
−2% · HOUSE EDGE0+3% · YOUR EDGE
0running count
6.0decks remaining
0.0true count
−0.5%player edge, roughly
2–6 COUNT +1 · 7–9 COUNT 0 · 10–A COUNT −1
Deal cards and watch the count. The edge is a function of the cards still in the shoe, nothing else. Shuffle, and everything you knew is gone. That is the Markov property at a card table.
That is a Markov process in its purest form. The state is the current shoe. The history of previous shoes is irrelevant by construction. And it is exactly why continuous shuffle machines kill card counting: when every hand is dealt from a fresh, full shoe, there is no state to track and no edge to find.
The parallel for a trader is direct. A market regime is a shoe. The recent tape (the last weeks or months of volatility, correlation, flow and participant behaviour) is the running count. A regime change is the shuffle. When the shuffle happens, your 15-year count is not just stale, it is actively misleading, because it tells you the shoe is rich when you are actually looking at a brand new deck.
Sports bettors operate the same way. Nobody prices a Sunday game off a team’s results from 2013. They price it off the current roster, current form, current injuries. The useful lookback is measured in weeks. Markets are harder than sport because the roster changes invisibly, but the logic of where the information lives is the same.
08 / THE HONEST COUNTERARGUMENT
What long history is actually for.
None of this means you should delete your 15-year chart. It means you should stop asking it the wrong question.
A short window has one real weakness: it has never seen a crash. Six months of calm data will tell you, with great confidence, that a 5 percent gap down cannot happen. That is not a forecast, it is a sample too small to contain the tail. NYU’s V-Lab says this plainly: GARCH-style models forecast well in normal conditions and cannot anticipate black swans or structural breaks. The same holds for any short-lookback method, including the Markov approach. Regime models tell you which state you are in. They are much worse at telling you when the state is about to flip.
So use the two histories for two different jobs:
| Question | Lookback | Why |
|---|
| Which regime am I in right now? | Weeks to 6 months | The Markov state lives in recent data; older data is a different state |
| What volatility should I expect tomorrow? | Days to weeks | Volatility clusters; the half-life of the signal is about 11 trading days |
| Does this specific setup have an edge today? | 3 to 12 months | Edges decay and participants change; a long window averages dead edges with live ones |
| How bad can it get? | 10 to 20 years or more | Tails are rare; you need the 2008s, the 2020s and the flash crashes to size them |
| How much should I risk per trade? | Both | Short window sets expected move; long window sets the worst case that decides survival |
In other words: trade the last six months, size for the last twenty years.
Long history belongs in your risk limits and your stop placement, not in your entry signals. The day traders who blow up are rarely the ones who used a short lookback. They are the ones who used a short lookback for everything, including the question of how much they could lose.
09 / THE ATOMICTRADES RULE OF THUMB
No longer than the regime, no shorter than you need.
If the market you are trading in is non-stationary, approximately Markov, volatility-clustered, populated by a new crowd, and subject to edge decay (and it is all five), then the most recent history is the most accurate history for every decision except tail sizing.
Our working rule
- Your lookback should be no longer than the current regime, and no shorter than you need to estimate it. For most intraday and swing setups that lands between six weeks and six months.
- When the tape changes character, treat it as a shuffled shoe. Volatility doubles, correlations flip, the flow that used to drive your setup disappears: reset the count.
- Do not reach for the 15-year chart to reassure yourself that the pattern will come back. The chart does not know the crowd changed.
- This is how the lab is built. It tests three, six and twelve months, tunes on the first two thirds of each window and judges on the last third the tuning never saw. Why we keep the history short ↗
The traders who last are not the ones with the longest backtests. They are the ones who notice the regime first and stop trading the old one.
10 / SOURCES
Where this comes from.
- Capponi, Huang, Sidaoui, Wang and Zou (2025), The Nonstationarity-Complexity Tradeoff in Return Prediction, arXiv 2512.23596
- Jebri (2014), Risk Analysis of Hedge Funds: A Markov Switching Model Analysis, Journal of Applied Business Research
- Finance Train, Models of Volatility Clustering: EWMA and GARCH(1,1) (RiskMetrics decay factor 0.94)
- NYU Stern V-Lab, Volatility Summary, vlab.stern.nyu.edu
- SEC Division of Trading and Markets (April 2026), Roundtable on Options Market Structure: Supporting Data
- Nunes (October 2025), 3 Warning Signs the VIX Won’t Tell You About Anymore, Investing.com
- Lee (2025), Not All Factors Crowd Equally: Modeling, Measuring, and Trading on Alpha Decay, arXiv 2512.11913 (includes the McLean and Pontiff 2016 figures)
- Liu (2026), Evaluating Structured Strategy Backtests: Peer Benchmarks, Regime Timing, and Live Performance, arXiv 2604.18821
- How to Count Cards: The Hi-Lo System and How to Count Cards in Blackjack
Hamilton’s 1989 regime-switching model (Econometrica) is cited from general knowledge; the table of RiskMetrics weights is computed from the 0.94 decay factor (0.94 raised to the number of trading days elapsed). The figures on this page are illustrations built from the numbers in the text, not live data. Nothing here is investment advice.