Thoughts on Curve Fitting
"Curve fitting (= overfitting) is not good"
This is in the EA industry,
as something that is assumed to be obviously true,
I think it has been spoken for many years,and that is what I believe.
Also, among some people,
"A high PF" = "because of curve fitting, forward performance cannot be expected"
and they may treat the two as synonymous.
However,
are "high PF" and "curve fitting"
the same thing (identical)?
In reality, up to now,
the EA that was thought not to be curve fit
has subsequently maintained an upward-facing forward performance?
When asked, unfortunately, most EAs
do not actually maintain an upward trend
in reality.
From this, we can see that
Ultimately,
the forward performance
isn't really related to whether the PF in backtests is high or low;
it simply depends on
whether the EA's logic has an edge
or not.
That, we believe, is the reality.
In other words,
if there is an edge in the underlying logic of the EA,
whether the backtest PF is low or high,
the forward performance will naturally be good.
And,
in cases where the logic has an "overwhelming edge,"
it will tend to have high PF in backtests as well,
and likewise, forward PF should also tend to be high,
that is the thinking.
If this is the case,
for a logic that originally had no edge,
forcing it to achieve a high PF would result in
an EA that is simply curve-fitted,
a mere "EA with no edge."
(※Such an EA is expected to turn downward immediately in forward testing)
In this sense,
for example, in my offerings, the
High PF EA Series,
(1) 'EA_final_max_Revolution' ⇒ (Backtest PF = ∞, Forward PF = ∞)
(2) 'EA_final_max_Revo_turbo' ⇒ (Backtest PF = 5.25, Forward PF = 3.78)
(3) 'EA_final_max_ED' ⇒ (Backtest PF = 1.95 (initial version PF 7.0), Forward PF = 3.51)
(Forward performance as of 2017/8/25)
As above, these EAs have achieved high PF in long-term backtests and
also maintained similarly high PF in a certain period of forward testing.
(For detailed performance, please see the sales pages for each EA below)
<＜List of Nekobyou's Listed EAs＞
https://fx-on.com/user/?i=1962&t=2
Thus,
the similarity between the backtest results and forward results
suggests that the logic has an edge,
and
it is highly likely that similar results can be expected in future EA operations.
(Of course, there is no guarantee that it will work forever, but at least
the EA's lifespan should accumulate profits for a period of time, I believe.)
In conclusion,
"Whether PF is high or low" and "whether curve fitting is present" do not necessarily coincide,
and the forward period performance becoming extremely poor is
due to an EA that has no edge, but merely curve-fitted.
And,
for EAs that have an edge in their logic
the PF in backtests tends to be high,
and due to that edge,
forward performance may be maintained similarly (i.e., high PF).
This is how I view it. Thank you.
※The above is solely the author's personal view and
does not guarantee accuracy or validity.
Neko Hakase
