Shelton [1997] describes
trading from an interesting perspective: Game Theory where each market investor
is calibrating his/her trades according to the different expected level of
volatility.
Lesson 101 in trading, states
to set stop-losses to trading positions. This is particularly true for
directional positions on cash or futures, where the downside risk can
potentially be unlimited.
Calibrating a stop-loss
implies to forecast future volatility, and therefore estimate future adverse
movements.
This can be done through an
ARMA on Integrated Volatility, or EWMA on Daily close-to-close returns
volatility for instance.
This is the linear component
of the future volatility assessment. Non-parametric, quite precise, good job!
Then, the question arises: can
we do better by integrating a non-linear component?
Let’s take a (locally)
trending market. Prices will oscillate within a 2 standard deviation from the
trend.
Pattern recognition implies to
search for any geometrical pattern drawn from extrema of these fluctuations
range.
One can think about this
problem as looking for patterns in a cloud wandering above our heads. No doubt
that thanks to human imagination, we will collect a very large variety of
answers. In financial charts, this is basically the same result. The complexity
in prices is so large that we will get very diverse answers. Coordination among
market participants usually can’t emerge from the ups and downs of the market
price (its volatility), because it’s completely random.
The easiest way to start
thinking consensus building is from local extrema of a range. From 2 local tops
with some distance, one can easily draw a trendline (upward or downward), and
try to infer from the subsequent price (re-)actions, if this trendline attracts
some attention.
Imagine the case of bullish
market configuration, with 3 aligned points (Graph 1) : S1 and S2 are support
points to draw a line a “V” acts as a validation point of this pattern.
In this case, after “V” being
confirmed, it is common-knowledge among market participants that the market is
coordinating along this trendline, and that therefore one shall expect some
significant buying orders, just before the price hit this trendline.
Well, if this is
common-knowledge this feature in the volatility estimate, why not integrate
this point in the trade calibration, as a non-linear component.
Funny to see that trendlines
are used to determine non-parametric & non-linear component of this
volatility forecast!

Price Action | Large Risk | Small Risk |
No Downside | w | w |
Small Downside | w | -x |
Large Downside | -y | -x |
Price Action | Large Risk | Small Risk |
No Downside | w | w |
Small Downside | w | -x |
Large Downside | -y | -x |
Non-Cooperative, Simultaneous Game between a trader and the market when consensus has emerged
In this special case, the trade calibration is altered by the consensus being detected. If the trader’s opinion is based solely on fundamentals, he/she will post a stop-loss at a a certain price level with a uniform probability. If the trader’s conviction is more price action-dependent, then he will be tempted to post his/her stop-loss immediately below the upward resistance trendline. Why ? Because he expects the trends to go on, and in case of adverse price movement, he is convinced that some buying limit orders will be posted just before this trendline.
If many traders post stop-losses below this trendline, in line with their own detection of a growing consensus building, one can expect a cascade of stop-losses if, inadvertently, the market price breaks the trendline and hit the stop-losses.

In this case, this will lead to a very particular price dynamics: an intraday “jump”, generated by this cascade of stop-losses. Jump in this situation is the signature of a past consensus, and contemporaneous unwinding of positions. Traders will then experience some slippage on their stop-losses, because of the local market illiquidity. Intraday Jumps in this case are said to be “UNEXPECTED”.