Forward Return Analysis: The Rigorous Way to Know If a Pattern Actually Predicts Anything
Last updated: March 2026
An event study is a statistical method that isolates a specific condition in market data and measures its impact on future returns. The idea is straightforward: if a pattern has predictive power, the returns that follow it should be measurably different from returns on an average day.
The methodology was developed in academic finance during the 1960s and 1970s, originally to study the impact of corporate events on stock prices. Researchers wanted to know: when a company announces a merger, what happens to the stock over the following days and weeks? When earnings beat expectations, how much of the move happens on day one versus day five versus day twenty?
The foundational insight was that you could not just look at raw price changes. You needed to measure abnormal returns — the difference between what actually happened and what would have happened if the event had not occurred. This requires a baseline, a statistical test, and a framework for handling multiple events that might overlap or cluster.
Systematic traders adapted this methodology for a different purpose. Instead of studying corporate announcements, they study technical and quantitative conditions: RSI dropping below a threshold, a moving average crossover, a volatility squeeze, a volume spike on a gap down. The question is the same. After this condition occurs, are future returns statistically different from normal?
The process has four steps, each of which matters:
The "event" is a boolean condition — something that is either true or false on any given bar. It could be simple (RSI(14) < 30) or complex (a multi-condition formula involving price relative to moving averages, volume ratios, and volatility measures). What matters is that the condition is unambiguous and can be evaluated on historical data without lookahead bias.
The quality of the event definition determines everything downstream. A vague event ("price looks oversold") is untestable. A precise event (RSI(14) < 30 AND close > SMA(200) AND ATR(14)/close < 0.02) is a concrete, falsifiable hypothesis.
Once the condition is defined, you scan the historical data and flag every bar where it evaluated to true. These are your "event dates." The number of events matters enormously. Five events prove nothing — the sample is too small for statistical inference. Fifty events start to become interesting. Two hundred events give you a dataset with real statistical power.
This is also where you confront practical questions. If the condition fires on three consecutive bars, is that one event or three? How you handle clustered signals affects your sample size and the independence assumption underlying your statistical tests.
For each event date, you measure what happened after the signal. This is forward return analysis. You do not look at what the market was doing before the signal or during the signal. You look exclusively at what happened next.
The critical design choice is measuring at multiple horizons. A standard set might be 1, 3, 5, 10, and 20 bars forward. For each event instance, you record the return at each horizon. This gives you a distribution of returns at each time scale.
Why multiple horizons? Because the same pattern can have radically different behavior at different time scales:
Testing a single horizon gives you one data point. Testing five horizons gives you the complete return profile of the pattern — when the edge appears, when it peaks, and when it fades. This shapes every downstream decision: how long to hold, where to place targets, and whether the pattern is worth trading at all.
The final step is the one that separates event studies from napkin math. You have an average forward return at each horizon. The question is not whether the average is positive or negative. The question is whether it is statistically distinguishable from zero.
An average 5-bar return of +0.8% means nothing by itself. If the standard deviation of those returns is 4%, you have a t-statistic of about 0.2 with most sample sizes — indistinguishable from noise. But if the standard deviation is 0.5%, the same average becomes highly significant. The dispersion matters as much as the mean.
Proper event studies run two significance tests at each horizon:
Both tests matter. A signal that passes both has genuine predictive power. A signal that passes only the first test might just be capturing the market's background drift.
Raw percentage returns have a fundamental problem: they are not comparable across different market environments. A 1% daily move in the S&P 500 during a quiet summer week is a significant event. A 1% move during a VIX spike is unremarkable. If your event study spans five years of data, some signals fired during calm markets and some fired during crises. Averaging raw returns across these regimes obscures what is actually happening.
ATR (Average True Range) normalization solves this by expressing returns in units of contemporaneous volatility. Instead of saying "the average forward return was +1.2%," you say "the average forward return was +0.7 ATR." This means the market moved 0.7 times its recent daily range in the expected direction — regardless of whether that daily range was 0.3% or 3%.
The benefits are substantial:
Event studies and backtests are not competing methodologies. They answer different questions, and the strongest research uses both.
| Dimension | Event Study | Backtest |
|---|---|---|
| Core question | Does this pattern predict returns? | Does this strategy make money with real execution? |
| Risk management | None (raw forward returns) | Stop loss, take profit, position sizing |
| What you learn | Predictive power of the signal itself | P&L, drawdown, Sharpe of a tradeable strategy |
| Horizons | Multiple (1, 3, 5, 10, 20 bars) | Variable (trade-by-trade, exit-dependent) |
| Overfitting risk | Lower (fewer free parameters) | Higher (SL, TP, hold time are all tunable) |
| Best for | Initial validation: is there anything here? | Strategy design: can I capture it? |
The natural workflow is to run the event study first. If the pattern shows no statistically significant forward returns at any horizon, there is nothing to capture — no stop-loss optimization, no entry timing adjustment, and no clever position sizing will create an edge from a pattern that does not predict returns. The event study is the gatekeeper.
If the event study does show an edge, the backtest determines whether that edge survives the friction of real trading. A pattern with a beautiful 5-bar forward return profile might still lose money after accounting for slippage, the cost of getting stopped out on false signals, and the opportunity cost of holding through drawdowns.
An event study tells you if the fish are in the lake. A backtest tells you if your rod, line, and technique can actually catch them.
Forward returns are easy to calculate. =INDEX(Close, ROW()+5) / Close - 1 gives you a 5-bar forward return. So why not just do this in Excel?
You can compute the numbers. What you will miss is everything that determines whether those numbers mean anything:
None of these problems are exotic. They are fundamental to any honest forward return analysis. A spreadsheet computes numbers. An event study framework computes answers.
Not all statistically significant results are equally useful. The two significance tests (versus zero, versus market baseline) create a natural four-category classification:
"No edge" is not a failure. It is the most common honest answer in quantitative research, and it is the answer that prevents the most damage. A trader who discovers that their favorite pattern does not predict returns has learned something genuinely useful. Most traders never run this test at all — they go straight to backtesting, optimize until the equity curve looks good, and then discover the hard way with real money that the pattern was noise.
One of the strongest forms of validation in event study research is testing the same pattern across multiple markets. The logic is straightforward: a pattern that captures a real market phenomenon should work on more than one instrument. An RSI oversold reversal that works on the S&P 500 but fails on the Nasdaq, the Dow, and the Russell is probably not capturing a universal oversold dynamic — it is capturing something idiosyncratic to the S&P's specific return distribution over the test period.
Conversely, a pattern that shows a strong edge on five equity index futures and no edge on gold and crude oil tells you something valuable about the pattern's scope. It works in equities specifically. That is useful information for portfolio construction and risk management.
Multi-market testing also functions as a partial guard against overfitting. A pattern that was accidentally curve-fit to one market's noise is unlikely to simultaneously fit the noise of four other uncorrelated markets. If it shows significance across all of them, you have much stronger evidence of a real phenomenon.
VARRD runs multi-market event studies from a single command, testing the same pattern formula across 2 to 10 markets in parallel. Each market gets its own significance tests, its own forward return profile at every horizon, and its own edge verdict. The results are then ranked by strength so you can see which markets respond most powerfully to the pattern and which are indifferent.
The traditional event study workflow requires a researcher who can write code, manipulate dataframes, implement statistical tests, handle edge cases around overlapping signals, normalize by volatility, and correct for multiple comparisons. This is not rocket science, but it is a multi-hour process for each pattern — and most people with trading domain knowledge do not have this specific skill set.
VARRD collapses this into a single natural language interaction. You describe your trading idea in plain English — "test whether RSI below 30 with price above the 200 SMA predicts a bounce on ES futures" — and the system handles everything:
The entire process — from idea to statistically validated result — takes about thirty seconds. If the event study shows an edge and you want to go further, you can run a backtest with stops, optimize stop-loss and take-profit parameters, or test the same pattern across additional markets. Each step builds on the last.
VARRD is accessible as a web application, through the MCP protocol at app.varrd.com/mcp for AI agents and tools like Claude Desktop and Cursor, and as a CLI for developers (pip install varrd). The event study engine is the same regardless of how you access it.
Most trading research tools are optimized to produce a "yes." Backtesting platforms let you tweak parameters until the equity curve points up. Optimizers find the stop-loss and take-profit combination that maximizes profit factor on historical data. The implicit assumption is that finding a profitable result is the goal.
Event studies invert this. The goal is not to find a profitable pattern. The goal is to know whether a pattern is profitable. That is a fundamentally different objective, and it leads to a different relationship with "no edge" results.
When an event study returns no statistically significant forward returns, the correct response is not disappointment. It is gratitude. You just learned — in thirty seconds, with rigorous statistics — that a pattern you might have spent months trading does not predict returns. Every dollar you do not lose on a non-edge is a dollar available for a real one.
The traders who build durable, long-term profitability are not the ones who find the most patterns. They are the ones who most efficiently discard the patterns that do not work.
What is an event study in trading?
An event study is a statistical method that measures what happens to asset returns after a specific condition or pattern occurs. You define the event (a boolean condition on price, volume, or indicator data), identify every historical instance, and measure forward returns at multiple horizons with statistical significance testing. The method originated in academic finance for corporate events and has been adapted for systematic trading research.
How is an event study different from a backtest?
An event study measures raw forward returns after a signal fires — no stops, no targets, no position sizing. It answers "does this pattern predict returns?" A backtest simulates an actual trading strategy with entry rules, stop losses, take profits, and execution constraints. It answers "does this strategy make money?" They are complementary: use the event study first to validate the signal, then backtest to design the strategy around it.
Why test at multiple time horizons?
The same pattern can behave very differently at different time scales. An oversold signal might show a strong 3-bar bounce but negative 20-bar returns. A momentum signal might show nothing at 1 bar but a significant move at 10-20 bars. Testing at 1, 3, 5, 10, and 20 bars reveals the full return profile — when the edge appears, when it peaks, and when it fades — which shapes hold period, target placement, and whether the pattern is tradeable at all.
What is ATR normalization and why does it matter?
ATR normalization expresses forward returns in units of recent volatility rather than raw percentage points. A 1% move during low volatility is very different from a 1% move during a crisis. Normalizing by ATR makes events from different volatility regimes and different markets directly comparable, and translates results into actionable stop and target distances.
How many signal occurrences do I need for a reliable event study?
More is better, but as a rough guideline: fewer than 20 events makes statistical inference unreliable. 30-50 events is the minimum for basic significance testing. 100+ events gives you meaningful statistical power and the ability to detect moderate effect sizes. If your pattern fires only 5 times in 10 years of data, the result — whether positive or negative — will have wide confidence intervals and limited reliability.
What does "no edge" mean in an event study?
"No edge" means the forward returns after the signal are not statistically distinguishable from zero or from the market baseline at any horizon tested. The pattern does not predict returns. This is a complete and valuable result — it tells you not to build a strategy around this pattern, saving you from losses on a non-edge. Most patterns tested honestly return "no edge," and accepting that is a core part of rigorous trading research.
Can I use VARRD's event study from my own code or AI agent?
Yes. VARRD exposes its event study engine through MCP (Model Context Protocol) at app.varrd.com/mcp, which is compatible with Claude Desktop, Cursor, and any MCP client. Developers can also use the Python CLI (pip install varrd) or the REST API. The event study runs the same statistical engine regardless of the access method.
Describe your idea in plain English. Get forward returns, p-values, and an edge verdict.
Event study, backtest, multi-market — one platform, rigorous statistics.
$2 free credits on signup. ~$0.30 per research session.
MCP: app.varrd.com/mcp | CLI: pip install varrd
This guide is maintained by VARRD Inc. and reflects VARRD's approach to event study methodology in systematic trading research. Last updated March 2026.