Why we keep
the history short.

01 / OUR THESIS

Two beliefs, one root.

At AtomicTrades we hold two beliefs that cut against most retail trading culture. First, day trading is gambling dressed up as a profession. Second, the most recent market history matters far more than a long chart stretching back years.

These two ideas come from the same root. Markets are not a fixed machine whose rules you can learn once from a decade of candles. They are a process whose rules keep changing, and the only reliable information about the next move lives in the current state of the market, not in its distant past.

That is the principle behind the Markov methods that quantitative hedge funds have used for decades: the future depends on where the process is now, not on the full path that led there. We build on that principle. We trade the regime we are in, we size for the risk that exists today, and we refuse to confuse frequent activity with edge.

02 / WHY DAY TRADING IS GAMBLING

Negative before you press buy.

Day trading loses money for almost everyone who does it persistently, and the data says so. The largest clean study, Day Trading for a Living? (Chague, De-Losso and Giovannetti, 2020), tracked every individual who started day trading Brazilian index futures between 2013 and 2015.

The authors also found no evidence that day traders improve with practice. The reason is structural, not a matter of discipline. Three forces turn a short-horizon trade into a bet with a negative expected value:

• Noise dominates signal at short horizons. On a five-minute chart, almost all price movement is random order flow. Whatever small predictable component exists is swamped by variance, so the outcome of any single trade is close to a coin flip.
• Costs are paid every flip. Spreads, fees, funding and slippage are charged on every entry and exit. A coin flip with a house fee on each toss is the definition of a casino game. Leverage multiplies the fee, not the edge.
• The counterparty is a machine. Intraday liquidity is provided by market makers and high-frequency firms with faster data, co-located servers and a structural advantage on every tick. A retail trader competing on their timeframe is the slowest player at the table.

What makes it feel like skill is the same thing that makes slot machines work: variable rewards. Winning streaks happen by chance, get attributed to talent, and the losses that follow get attributed to bad luck. The screen time, the adrenaline and the sense of control are the product being sold. The returns are not.

We are not against trading. We are against trading on timeframes where the expected value is negative before you even press buy.

03 / WHY LONG HISTORIC CHARTS MISLEAD

Ten years of chart, several markets.

A ten-year chart is not ten years of evidence about today’s market. It is a stitched-together record of several different markets that happened to trade under the same ticker.

The participants change. Bitcoin in 2014 was traded by hobbyists on a handful of exchanges. Bitcoin in 2026 is traded by ETFs, market makers, treasury companies and perpetual futures venues with billions in open interest. The price series looks continuous, but the process generating it was replaced several times along the way.

The mechanics change. Liquidity, leverage limits, funding rates, correlation to equities, the share of volume that is algorithmic, and the macro backdrop all move. A support level from 2019 means nothing to a market maker’s inventory model in 2026. Back-testing a rule across the whole history mostly tells you how it performed in markets that no longer exist.

The statistics change. Volatility clusters: calm periods stay calm, violent periods stay violent, and the switch between them is abrupt. An average computed over a decade blends a bull run, a crash, a sideways grind and a liquidity crisis into one meaningless number. The volatility of the last few weeks predicts the volatility of the next few weeks far better than the volatility of the last few years does.

This is why we treat the distant chart as context at best and as a trap at worst. The more data you add from the far past, the more you dilute the one thing that actually carries information: where the market is right now and how it has been behaving lately.

04 / THE MARKOV PRINCIPLE

Given where you are, the path adds nothing.

A Markov process is one where the probability of the next state depends only on the current state, not on the path that led to it. Mathematicians call this the memoryless property. In plain terms: given where you are now, the history of how you got there adds nothing.

Applied to markets, this is the formal version of our second belief. If prices are close to Markovian, then a long chart is not a long memory. It is a long record of states the market has already left. The only state that constrains tomorrow is today’s.

The practical version used in finance is the hidden Markov model. The market sits in one of a small number of unobservable regimes (for example quiet uptrend, high-volatility selloff, sideways chop). We never see the regime directly; we see noisy prices that it emits. Each regime has its own typical return and volatility, and a transition matrix gives the odds of switching from one regime to another. Leonard Baum, who later became Jim Simons’ first partner at Renaissance Technologies, co-developed the Baum-Welch algorithm that lets a computer infer these hidden regimes and transition odds from price data alone.

Two consequences follow, and they are the heart of the AtomicTrades method:

1. Identify the current regime, then trade its statistics. The question is never “what did this coin do in 2021?” but “which regime are we in now, and what does this regime usually do next?”
2. Weight recent data heavily. The most informative observations about the current state are the newest ones. Older data mostly describes other regimes, so it gets discounted, not averaged in.

This also explains why we reject intraday trading from a different angle. At very short horizons, the regime does not change from one bar to the next, so there is no state transition to trade. All that is left is noise around a fixed state. The edge in a Markov view lives in correctly reading the regime and positioning for its persistence or its switch, which plays out over days and weeks, not minutes.

05 / HOW HEDGE FUNDS USE IT

The best record in the business reads the present.

The most profitable trading operation in history was built on exactly this idea. Jim Simons spent his early career at the Institute for Defense Analyses breaking codes, which is the job of finding hidden states behind noisy signals. When he started trading in 1978, he recruited Leonard Baum, whose Baum-Welch algorithm trains hidden Markov models. Their approach treated market prices as the visible output of an unobservable process and tried to infer that process from recent data rather than from economic narratives. That lineage runs straight into Renaissance’s Medallion Fund, launched in 1988, which later folded hidden Markov models, kernel methods and other pattern-recognition tools into a single monolithic model (Zuckerman, The Man Who Solved the Market).

The same family of tools is now standard across systematic funds under the name regime-switching or Markov-switching models. A Cambridge study fitting a hidden Markov model to futures returns found that two or three hidden states, estimated with Baum-Welch, were enough to describe the data and generate tradable momentum signals. A two-state model on one equity series explained about 65 percent of the variation in returns, with the market spending most of its time in a calm state and switching to a stressed state roughly once a month (Kinlay).

What these funds have in common, and what retail day traders lack:

The lesson is not that retail traders should build their own hidden Markov model. It is that the people with the best track record in the business treat the market as a state machine driven by its present condition, and they treat long history as raw material for estimating transitions, not as a map of the future.

06 / WHAT THIS MEANS AT ATOMICTRADES

Stop asking what the chart did. Ask what the market is doing now.

Everything we build follows from the two beliefs above. In practice that means:

• We trade regimes, not candles. Our first job on any market is to classify the current state from recent price, volatility and funding behaviour. Position direction and size follow from the regime, not from a pattern on a chart.
• Recent history is weighted, old history is discounted. Our models give the last few weeks far more influence than the last few years. Older data is used to estimate how regimes transition, never as a direct forecast.
• Our horizon is days to weeks. That is where regime persistence and regime switches play out, and where the edge is large relative to costs. We do not scalp, and we do not hold through regimes we no longer understand.
• Volatility sets size. Leverage is a function of current volatility and the distance to invalidation, not of confidence. The same idea with twice the volatility gets half the size.
• We expect to be wrong often. A regime read is a probability, not a certainty. The method works because the wins in correctly identified regimes outweigh the losses from misreads, not because any single trade is safe.
• Fewer, larger, better. If a trade needs to be taken right now or the opportunity is gone, it is probably noise. Real regime shifts give you time to act.

The short version we tell new members: stop asking what the chart did. Ask what the market is doing now, and how long that usually lasts.

07 / HOW THE INDICATOR LAB APPLIES IT

Twelve months, and not a day more.

The lab on our home page is this philosophy turned into a machine. Every choice in it comes from the two beliefs above, and the first choice is the one people ask about most: why we never test more than twelve months.

Why the lab still lets you test fast charts

The lab runs on 15-minute, hourly, 2-hour, 4-hour and daily bars. We keep the fast ones in, not because we recommend them, but because the quickest way to learn why we don’t is to watch it yourself: every backtest charges a trading cost on every entry and exit, and the verdict is read on bars the tuning never saw. A combination that looks brilliant on the tuned two thirds of a 15-minute chart usually comes apart on the unseen third once every flip has paid its spread. You will see the score fall as the chart gets faster and the trades get more frequent. That is the point.

Our own forex rules work the 1-hour chart, one trade at a time, inside the London morning, with a stop and a target decided when the bar closes: a handful of trades a week, not hundreds of flips a day. That is as short as we go, and the scanner keeps every losing signal on the record so you can judge it for yourself.

Where the rest of the method shows up

• Regime, in plain sight. The market context strip reads sentiment gauges and a news agent per market, as a second opinion next to the lab’s signals, never as a signal on its own.
• Nothing repaints. Every signal is decided when its bar closes and never moves afterwards. We do not take that on trust: every indicator in the library, and every one the builder finds, is run on a full series and on the same series cut short, and rejected if any signal on the shared bars differs.
• Streaks, not averages. Every result shows its longest winning and losing run and its worst drawdown, because a regime read is a probability and you will sit through the misreads.
• Volatility in every stop. Stops and targets are set in multiples of the average true range of the last fourteen bars, so the same idea gets a wider stop and a smaller size when the market is wilder.
• Free data, said plainly. Where a free feed has holes, another fills them and the result tells you which weeks. A score is a reason to look closer, not a promise.

08 / READING LIST AND SOURCES

Where this comes from.

• Gregory Zuckerman, The Man Who Solved the Market (2019): the Simons, Baum and Medallion history.
• Fernando Chague, Rodrigo De-Losso and Bruno Giovannetti, Day Trading for a Living? (2020): the Brazilian day trader study.
• Hugh Christensen, Richard Turner and Simon Godsill, Hidden Markov Models Applied to Intraday Momentum Trading (University of Cambridge): a worked regime model on futures.
• Jonathan Kinlay, Regime-Switching and Market State Modeling: a two-state Markov fit with transition probabilities.

Nothing here is investment advice. It is a statement of the beliefs behind how we trade and how the lab is built.