Add logarithmic interpolation predection

Slightly better than linear regression.

2 files changed, 23 insertions, 0 deletions

diff --git a/log_regress_pred.m b/log_regress_pred.m
new file mode 100644
index 0000000..3c90441
--- /dev/null
+++ b/log_regress_pred.m
@@ -0,0 +1,22 @@
+function pred = log_regress_pred(price, quantity)
+
+	days = size(price, 1) - size(quantity, 1);
+	product_count = size(price, 2);
+	test_count = size(quantity, 1);
+	test_set = 1:test_count;
+	pred_set = test_count+1:size(price, 1);
+
+	pred = zeros(days, product_count);
+	for i = 1:product_count
+		[a,b] = regress_log(price(test_set, i), quantity(test_set, i));
+		pred(:, i) = a * log(price(pred_set, i)) + b;
+	end
+
+endfunction
+
+
+function [a,b] = regress_log(x, y)
+	ab = [ log(x) ones(length(x), 1) ] \ y;
+	a = ab(1);
+	b = ab(2);;
+endfunction
diff --git a/run_tests.m b/run_tests.m
index 6ad92ce..b7b1a83 100644
--- a/run_tests.m
+++ b/run_tests.m
@@ -6,6 +6,7 @@
 % plot prediction quallity
 calc_error('mean', q(29:42, :), mean_pred(q(1:28, :), 14));
 calc_error('regress', q(29:42, :), regress_pred(p, q(1:28, :)));
+calc_error('log regress', q(29:42, :), log_regress_pred(p, q(1:28, :)));
 calc_error('sevenday', q(29:42, :), repmat(sevenday_pred(q(1:28, :), 4), 2, 1));
 calc_error('random', q(29:42, :), rand_pred(q(1:28, :), 14));