Observation
While reviewing a recent dataset of 205 trading signals, I identified a subset of trades where the signal did not transition to flat (0) before reversing direction.
Instead, the signal immediately flipped:
1 → -1
-1 → 1
These trades were classified as flip trades.
At first, the spreadsheet logic only calculated PNL correctly when a signal transitioned from long or short into flat. After modifying the Excel formulas to properly handle reversal conditions, the hidden behavior of these flip trades became visible.
That visibility led to an important observation.
Dataset Results
Total Signal Population
| Metric | Value |
|---|---|
| Total Signals | 205 |
| Flip Trades | 30 |
| Non-Flip Trades | 175 |
Flip Trade Performance
| Metric | Value |
|---|---|
| Winning Flip Trades | 5 |
| Losing Flip Trades | 25 |
| Win Rate | 16.7% |
| Net Result | -81.5 ES Points |
| Average Expectancy | -2.72 Points Per Trade |
What Is a Flip Trade?
A flip trade occurs when the signal reverses direction immediately instead of first returning to flat.
Examples:
Long → Short
1 → -1
Short → Long
-1 → 1
In this trading framework:
- Signals are generated on bar close
- Trades enter on the next bar open
- The spreadsheet tracks both signal state and active position state separately
Because of that architecture, the PNL and exit logic required additional formula handling to correctly recognize reversal exits.
Once the spreadsheet logic was corrected, the reversal regime became measurable.
Initial Characteristics of Flip Trades
The flip trades showed several common characteristics:
- Short duration
- Weak directional follow-through
- Choppy market structure
- Rapid reversals
- High frequency of immediate failure
This strongly suggests that the flips are occurring during low-quality market conditions rather than during strong trend transitions.
Possible Explanations
Several hypotheses may explain the poor performance of reversal flips:
1. Noise Sensitivity
The model may be overreacting to short-term countertrend movement.
2. Failed Breakout Conditions
Some flips may occur after breakout attempts fail and reverse quickly.
3. Consolidation / Chop
The reversals may cluster during:
- low volatility
- overlapping moving averages
- lunchtime conditions
- compressed ranges
- mean reversion environments
4. Delayed Recognition
Some reversals may represent the model recognizing a losing trade too late.
Spreadsheet Research Improvements
This observation only became visible after modifying the Excel formulas responsible for:
- trade exit detection
- reversal recognition
- PNL calculation
The updated spreadsheet logic now properly handles:
1 → 0
-1 → 0
1 → -1
-1 → 1
That change transformed hidden spreadsheet behavior into measurable trading intelligence.
New Analytics Added
The next step is separating flip trades from normal trades for statistical analysis.
Flip Detection Formula
=IF(ABS(CurrentSignal-PreviousSignal)=2,"FLIP","NORMAL")
This allows:
- separate expectancy calculations
- separate win rate tracking
- duration analysis
- MFE / MAE comparison
- volatility regime analysis
- time-of-day analysis
Current Operational Decision
For now, flip trades will not be taken live.
The current evidence suggests that reversal flips are:
- statistically weak
- low quality
- high noise
- expectancy negative
Additional research is required before these conditions can be trusted.
Future Research Directions
Several future experiments may improve reversal quality.
Potential Filters
- Random Forest confidence
- breakout strength confirmation
- ATR expansion
- EMA separation
- minimum trend duration
- cooldown periods between reversals
- volatility regime filters
Machine Learning Possibilities
One particularly interesting direction is building a specialized classifier:
Will this flip trade fail?
Instead of simply asking:
Should I take this signal?
The existing dataset already contains:
- labels
- outcomes
- durations
- excursions
- contextual indicators
which makes this an ideal candidate for future Random Forest experimentation.
Key Takeaway
Sometimes the most important discoveries are not found by adding complexity.
They are found by fixing the measurement system.
In this case, a spreadsheet formula update exposed a hidden trade regime that appears to be materially damaging overall expectancy.
That is exactly the type of observation-driven research process that Trade Observations was built to support.