1 00:00:00,000 --> 00:00:18,080 Well, good morning everybody. How are you doing this week? Good. Yeah, good. I want 2 00:00:18,080 --> 00:00:24,580 to thank Charles real quick. So Charles sent me the pipeline workbook he put together last 3 00:00:24,580 --> 00:00:36,680 night with waste by licensee by day, which is the exact data point that better carbon 4 00:00:36,680 --> 00:00:42,680 solutions has been searching for. And I think it'll be helpful for anyone who's interested 5 00:00:42,680 --> 00:00:49,240 in cannabis waste. And I think we were talking about last week, how perhaps people in other 6 00:00:49,240 --> 00:00:57,560 states could even look at Washington and learn from us. So it could be helpful to waste producers 7 00:00:57,560 --> 00:01:05,320 in other states as well. Good work there, Charles. Do you have any comments? So, you 8 00:01:05,320 --> 00:01:09,920 know, definitely take a look at it and see how you want to display that. There's over 9 00:01:09,920 --> 00:01:17,800 like 1100 different producers that produced waste. And so I wasn't sure how to like display 10 00:01:17,800 --> 00:01:31,200 that. Well, I would just start with just just average. So just what's the average amount 11 00:01:31,200 --> 00:01:41,840 of waste for licensee per day. And so I would maybe start there. So that that way you can 12 00:01:41,840 --> 00:01:48,680 just get a singular time series. You could look at like the top 10, you know, see what 13 00:01:48,680 --> 00:01:59,160 the top 10 producers how much waste they're producing. And then the this one, you didn't 14 00:01:59,160 --> 00:02:03,480 really need to sort it by licensee. But it would also just be interesting to look at 15 00:02:03,480 --> 00:02:15,240 total waste by day. So those are the three metrics that I even start by plotting. And 16 00:02:15,240 --> 00:02:21,840 then those are just all time series. So you could get more creative than that, I'm sure. 17 00:02:21,840 --> 00:02:32,160 Okay. I'll keep plugging away at it. Speaking of getting creative, if you were so one of 18 00:02:32,160 --> 00:02:38,760 my favorite types of figures is a map. So you could potentially have a map and have 19 00:02:38,760 --> 00:02:46,040 almost I'm not sure you could do it. I wonder if you could almost do like, like radial sizes. 20 00:02:46,040 --> 00:02:54,680 So you just have bigger circles around the bigger bigger waste producers. But I'm not 21 00:02:54,680 --> 00:03:00,980 sure if that would really make sense. That way you could almost get like a geographical 22 00:03:00,980 --> 00:03:08,200 sense of where all the waste is. Yeah, maybe like a heat map of the region. 23 00:03:08,200 --> 00:03:13,960 Exactly. A heat map would be a good one. If you're feeling ambitious, but you have the 24 00:03:13,960 --> 00:03:22,840 data so it's possible. And if you need to geocode the addresses, canlytics has a little 25 00:03:22,840 --> 00:03:29,280 utility helper that for the most part, you can just toss in someone's name, just their 26 00:03:29,280 --> 00:03:40,480 licensee name, and it can get you the geo coded address. But anywho, good work there. 27 00:03:40,480 --> 00:03:46,360 And I think how about maybe I'll take a look at it. We can digest it. And then if you want 28 00:03:46,360 --> 00:03:51,160 it, we could look at that to look at your waste data next week. 29 00:03:51,160 --> 00:03:52,160 Okay. 30 00:03:52,160 --> 00:04:05,440 So that's awesome. Good to see you here too, Nick. I've actually got something interesting 31 00:04:05,440 --> 00:04:14,720 for today. So last week we started looking at the inflation of cannabis prices in Oregon. 32 00:04:14,720 --> 00:04:28,720 So finished compiling that data. So let me just go ahead and push it up to the Canvas 33 00:04:28,720 --> 00:04:46,640 Data Science repository here. And then I will go ahead and start showing you just a bit 34 00:04:46,640 --> 00:05:03,720 of the progress that I made on calculating inflation rates since last week. Don't worry, 35 00:05:03,720 --> 00:05:14,560 Charles, didn't miss anything. I'm just opening up the presentation. 36 00:05:14,560 --> 00:05:31,240 All right. So it is actually April 21st. However, this is essentially the presentation I put 37 00:05:31,240 --> 00:05:37,240 together last week. So Charles has seen this. Nick, I'm just going to run through this real 38 00:05:37,240 --> 00:05:44,200 quick just so that you can get up to speed with the data we're looking at and the stats 39 00:05:44,200 --> 00:05:52,280 we're trying to calculate here. So we're looking at total cannabis sales in Oregon, and we 40 00:05:52,280 --> 00:06:01,400 can break that down by product type. And this is available through Oregon's public data 41 00:06:01,400 --> 00:06:11,480 dashboard for the cannabis industry. And they update their data quickly. So you have everything 42 00:06:11,480 --> 00:06:18,280 through the prior month. So you have everything up through March of 2021, which is phenomenal 43 00:06:18,280 --> 00:06:23,760 that they publish their data so quickly. Yeah. Did they fix that dashboard? Because last 44 00:06:23,760 --> 00:06:29,520 time I checked, it was like, I always had issues, but maybe I had this wrong link. But 45 00:06:29,520 --> 00:06:33,320 like it would always just air out and nothing would show up. So that's... 46 00:06:33,320 --> 00:06:46,480 Yes. So the dashboard is a little tricky. So when you access it, it's actually an HTTP 47 00:06:46,480 --> 00:06:56,880 address and you may have to hit the advance, proceed to an unsafe website button. Okay. 48 00:06:56,880 --> 00:07:07,720 All right. And so it may look like the website's unavailable at first, but essentially I think 49 00:07:07,720 --> 00:07:17,480 they're just running on HTTP. And then most modern browsers warn you against navigating 50 00:07:17,480 --> 00:07:23,200 to some HTTPs. It's worse than that. It means that they're 51 00:07:23,200 --> 00:07:28,200 running HTTPs, but they have a self-signed cert or one that can't be verified. 52 00:07:28,200 --> 00:07:38,880 Oh, thanks for chiming in. So exactly. So it's good to hear from someone who knows their 53 00:07:38,880 --> 00:07:49,440 computer science. So thank you for chiming in. So maybe it preceded your own caution, 54 00:07:49,440 --> 00:07:56,360 but in this case I went ahead and collected the data because the data is valuable in our 55 00:07:56,360 --> 00:07:59,360 case. Yeah. 56 00:07:59,360 --> 00:08:09,360 So the other thing is you may figure out a niftier way. I essentially... So here I'll 57 00:08:09,360 --> 00:08:26,880 show you what the data portal looks like. So as we've noticed, it's not secure, but 58 00:08:26,880 --> 00:08:37,920 we're not posting any sensitive data here. So long story short, you can download some 59 00:08:37,920 --> 00:08:49,920 series, but some series it's harder to download. So I actually just transcribed the figures. 60 00:08:49,920 --> 00:08:59,480 And I put them here in the... Well, I may have saved... Oh, yes. So essentially I've 61 00:08:59,480 --> 00:09:13,200 just transcribed the numbers here into Excel. So you can double check my work that I transcribed 62 00:09:13,200 --> 00:09:23,200 everything correctly. But if you can think of a niftier way to collect this data, then 63 00:09:23,200 --> 00:09:33,400 by all means I've added. So back to the presentation. We've collected sales data, we've collected 64 00:09:33,400 --> 00:09:45,680 price data, and we're curious about inflation of prices. So of course you notice right off 65 00:09:45,680 --> 00:09:54,360 the bat that prices in the first two years decreased dramatically as you would expect 66 00:09:54,360 --> 00:09:59,680 in a new market. So as things are coming to equilibrium, as people are figuring out how 67 00:09:59,680 --> 00:10:08,000 to price their products. And then you notice in the longer run, prices seem to stabilize 68 00:10:08,000 --> 00:10:18,120 and you have steady... It appears that you may have steady or inflation at this point. 69 00:10:18,120 --> 00:10:29,840 And so that would just be a moderate increase in prices over time. And then we've noted 70 00:10:29,840 --> 00:10:38,920 that economic theory would suggest that sales and prices would be correlated with the interest 71 00:10:38,920 --> 00:10:46,720 rate. We notice the cannabis industry is a little unique, so it may not be as strongly 72 00:10:46,720 --> 00:10:52,440 correlated with the interest rate as other industries may be. And so we're just going 73 00:10:52,440 --> 00:11:00,800 to see if the interest rate can be helpful to predict inflation in the cannabis industry 74 00:11:00,800 --> 00:11:13,280 or not. So here you see the interest rate is lower to near zero in early of 2020, March 75 00:11:13,280 --> 00:11:26,200 of 2020 to be specific. Well, maybe February. Next, we're going to try to estimate output, 76 00:11:26,200 --> 00:11:32,680 inflation and the interest rate. And just for a reasonable timeline, we'll just do through 77 00:11:32,680 --> 00:11:40,080 2022. So just for the remainder of the year to see what output and inflation may be in 78 00:11:40,080 --> 00:11:52,040 Oregon this year. So we are using economic theory. However, the statistical model is 79 00:11:52,040 --> 00:12:01,960 a theoretical. So it's really just taking any vector of time series and using historical 80 00:12:01,960 --> 00:12:12,920 observations to forecast that data forward. It's useful because you can apply it to really 81 00:12:12,920 --> 00:12:23,680 any time series data. And as long as it's coming in at a regular interval. So you can 82 00:12:23,680 --> 00:12:33,400 look at daily data. If you're looking at actual individual transactions during the day, there's 83 00:12:33,400 --> 00:12:40,600 statistical problems there. And then the other major pitfall of vector autoregressions is, 84 00:12:40,600 --> 00:12:50,120 of course, if you look at our equations from earlier, you'll see that there are a lot of 85 00:12:50,120 --> 00:13:04,480 variables. So sales depend on sales from however many periods ago. So they may depend on one, 86 00:13:04,480 --> 00:13:15,400 two, six periods ago. Likewise, it also depends on inflation from one to six periods ago, 87 00:13:15,400 --> 00:13:25,600 ten periods ago. However, many lags are in the model, similar for inflation. So if you 88 00:13:25,600 --> 00:13:35,320 only have, so in this case, we're looking at monthly price data. So we're just looking 89 00:13:35,320 --> 00:13:47,560 at the number of months really between 2017 and early of 2021. So we have about 50 observations 90 00:13:47,560 --> 00:13:58,420 here. And so we're going to be chewing through our degrees of freedom quite quickly. Because 91 00:13:58,420 --> 00:14:14,360 if, say, you just had one lag, then in this first model, you would have four parameters 92 00:14:14,360 --> 00:14:28,000 times three models. So that would be 12 parameters with just one lag order. So with 50 observations, 93 00:14:28,000 --> 00:14:36,440 you probably couldn't estimate the VAR with a lag order of six. You probably would not 94 00:14:36,440 --> 00:14:48,560 have enough observations. So keep that in mind when you're estimating your VARs. So 95 00:14:48,560 --> 00:14:57,480 that's why monthly data can be a little tricky. Weekly data is nice. And then daily data has 96 00:14:57,480 --> 00:15:10,400 its advantages because you get a lot of observations. And my only observation is it depends on your 97 00:15:10,400 --> 00:15:22,080 forecasting horizon of which frequency you want to choose. This is just a slight note 98 00:15:22,080 --> 00:15:30,440 that just says that we don't actually think that... Well, essentially, we're just saying 99 00:15:30,440 --> 00:15:36,200 we think that people essentially take the interest rate into their calculations when 100 00:15:36,200 --> 00:15:43,560 they're making decisions. So the Federal Reserve can change it however they want. People are 101 00:15:43,560 --> 00:15:53,440 just going to change prices and act accordingly. So enough of that. Let's go ahead and get 102 00:15:53,440 --> 00:16:04,880 into forecasting. So real quick, just to run over the 10. So these are essentially... Professor 103 00:16:04,880 --> 00:16:14,400 Iqbal, UNC Charlotte just told me these 10 steps to forecasting. And they've proved pretty 104 00:16:14,400 --> 00:16:20,760 useful as 10 steps to follow when you're doing forecasting. But first, we need to know what 105 00:16:20,760 --> 00:16:29,680 we're forecasting, the inflation rate in Oregon, cannabis prices. We need to understand why 106 00:16:29,680 --> 00:16:38,520 we're forecasting because we want to know how prices are going to be moving in Oregon. 107 00:16:38,520 --> 00:16:43,320 As we've saw, it looks like there's going to be moderate inflation, but we want to actually 108 00:16:43,320 --> 00:16:50,800 quantify that. And we need to acknowledge the cost of the forecast error. So that would 109 00:16:50,800 --> 00:16:59,640 just mean we can't take too much stock into our forecast. So what's the cost of over 110 00:16:59,640 --> 00:17:07,080 estimating inflation and what's the cost of underestimating inflation? Because if we put 111 00:17:07,080 --> 00:17:13,040 these forecasts out there and businesses make decisions about them based on what they think 112 00:17:13,040 --> 00:17:20,960 prices may be, they need to take into consideration what's the chance of overestimation slash 113 00:17:20,960 --> 00:17:28,000 underestimation. That's a tricky one, but important to take into consideration. 114 00:17:28,000 --> 00:17:34,200 Next, we just need to know the horizon, which we're just going to be forecasting through 115 00:17:34,200 --> 00:17:45,920 2022. Long enough to be useful, however, not too long that it's unrealistic. We've chosen 116 00:17:45,920 --> 00:18:01,920 our variables, inflation, output, and the interest rate. And we've chosen them because 117 00:18:01,920 --> 00:18:15,880 of economic theory. We've then chosen a vector autoregression to do the statistics. And the 118 00:18:15,880 --> 00:18:21,680 vector autoregression is a vector that fits our economic theory and it's a relatively 119 00:18:21,680 --> 00:18:28,000 simple statistical model that we can also use for forecasting. 120 00:18:28,000 --> 00:18:34,720 We'll want to present the results in some sort of figure. I'm not sure if you've heard 121 00:18:34,720 --> 00:18:42,000 of Edward Tuft, but he would always emphasize that you need to show the data in some sort 122 00:18:42,000 --> 00:18:52,680 of beautiful figure. So we'll throw out some charts. Then we'll want to interpret what 123 00:18:52,680 --> 00:18:58,440 predictions we've made. So we'll want to see if inflation is rising and decreasing, what's 124 00:18:58,440 --> 00:19:05,920 our confidence. Next, we will essentially want to use recursive 125 00:19:05,920 --> 00:19:13,760 methods. So what this means is we want to keep forecasting each month. So we'll want 126 00:19:13,760 --> 00:19:23,120 to revisit our forecasts next month and compare our forecasts to the actuals and see if we 127 00:19:23,120 --> 00:19:29,960 can't make better forecasts the next month. And when we do that, that brings us to the 128 00:19:29,960 --> 00:19:36,120 final step is when we look at our forecasts next month, so we'll be forecasting April 129 00:19:36,120 --> 00:19:43,400 sales today. And then in May, we can actually check our 130 00:19:43,400 --> 00:19:50,760 forecasts and see if the model we selected was a good predictor. And we may need to select 131 00:19:50,760 --> 00:19:58,960 a different model if our forecasts were wildly wrong or slightly adapt the model or slightly 132 00:19:58,960 --> 00:20:08,640 stick with it if we can't find a better model. So that's a quick background. Are there any 133 00:20:08,640 --> 00:20:27,320 questions before we jump into the code here? All right. So I just opened up Spyder here. 134 00:20:27,320 --> 00:20:35,560 There's nothing special about Spyder. I'm getting used to running Python in VS code. 135 00:20:35,560 --> 00:20:45,840 However, I'm still just faster and more effective in Spyder. So until I'm better in VS code, 136 00:20:45,840 --> 00:20:55,960 we'll be here. Or another editor. I've heard of PyCharm and Atom. Both seem like good choices. 137 00:20:55,960 --> 00:21:08,560 I haven't tried either. You're a weapon of choice. 138 00:21:08,560 --> 00:21:21,120 The packages we're using today are essentially NumPy and Pandas. And then I'll be reading 139 00:21:21,120 --> 00:21:32,080 in a Federal Thread API key from an environmental variable, an environment variable. And then 140 00:21:32,080 --> 00:21:46,240 we'll be using the Thread API to get the interest rate. 141 00:21:46,240 --> 00:21:56,120 First things first, we need to read in the data. So I'll just be reading in this data 142 00:21:56,120 --> 00:22:15,600 that I've scrounged from the Oregon dashboard. Just to look at the first observations. 143 00:22:15,600 --> 00:22:32,000 Notice, we don't really have consistent prices across the board until about 2017. So a professor 144 00:22:32,000 --> 00:22:39,320 in college really emphasized you never want to throw away data. However, in this case, 145 00:22:39,320 --> 00:22:47,200 just for simplicity's sake, I'm restricting the time period to 2017 and onwards. And that's 146 00:22:47,200 --> 00:22:52,520 for simplicity's sake. If you have a more elegant way to estimate the models, then by 147 00:22:52,520 --> 00:22:58,680 all means. 148 00:22:58,680 --> 00:23:11,320 We're looking at data from 2017 through March of 2021. And first we'll look at the data. 149 00:23:11,320 --> 00:23:23,280 So what are some of the variables we have here? We've got total sales. We've got flower 150 00:23:23,280 --> 00:23:51,440 sales. We've got concentrate sales. I thought we had retail flour prices. Okay, so we've 151 00:23:51,440 --> 00:24:04,920 got retail flour prices. And we have retail concentrate prices. Those are our variables 152 00:24:04,920 --> 00:24:08,920 of interest. 153 00:24:08,920 --> 00:24:19,680 And so last week, we talked briefly about how to calculate the CPI. And just to show 154 00:24:19,680 --> 00:24:35,080 you a presentation from last week with my chicken scratch. We basically noted that the 155 00:24:35,080 --> 00:24:47,840 consumer's basket of goods is about 60% flour and about 40% concentrate. So if you were 156 00:24:47,840 --> 00:24:56,800 going to calculate the CPI, which is a price index. So you can think about a stock index, 157 00:24:56,800 --> 00:25:05,000 which just sort of measures the general trend of all the goods in the basket. And so our 158 00:25:05,000 --> 00:25:18,760 basket is concentrates in flour. So our CPI will be the sum of all prices times their 159 00:25:18,760 --> 00:25:28,960 weight. So we can actually do that here in Python or your favorite programming language. 160 00:25:28,960 --> 00:25:47,880 So essentially, we can calculate the share of flour out of total sales. 161 00:25:47,880 --> 00:25:57,120 Rule number one, look at the data. So as you see, flour was making about 70% of all sales. 162 00:25:57,120 --> 00:26:07,880 And it's actually decreased, which is interesting, and is now maybe kind of stabilizing around 163 00:26:07,880 --> 00:26:20,120 55 or so percent. That's quite interesting. Then you can calculate concentrate share of 164 00:26:20,120 --> 00:26:31,640 the sales. And so this should be roughly the inverse where you have concentrates. Because 165 00:26:31,640 --> 00:26:39,720 keep in mind, there are also edibles and other goods. So we are excluding some data here. 166 00:26:39,720 --> 00:26:46,880 So it would be nice to have prices on edibles, but edibles vary so much across the board 167 00:26:46,880 --> 00:26:56,400 that it may not be key of pairing apples to apples, or infused apple to infused apple. 168 00:26:56,400 --> 00:27:01,960 Yeah, that's what sucks about this is they kind of lumped everything together, like edibles 169 00:27:01,960 --> 00:27:09,800 and like some of that. Because the concentrate would include oil then too, right? 170 00:27:09,800 --> 00:27:11,240 Yes. 171 00:27:11,240 --> 00:27:14,600 So it's a lot of different things that they put in there. And then on the manufacturing 172 00:27:14,600 --> 00:27:19,560 side, we have to consider this as pie retail sales. So if we're forecasting, we have to 173 00:27:19,560 --> 00:27:25,960 actually consider there's a markup. Because we're trying to get back to what we sell it 174 00:27:25,960 --> 00:27:32,840 at. So it gets a little complicated. 175 00:27:32,840 --> 00:27:41,080 That's a really good observation from someone who's in the industry. Because as you noted, 176 00:27:41,080 --> 00:27:48,200 you're sort of aggregating everything here. As you said, there's a wide array of concentrate 177 00:27:48,200 --> 00:27:56,280 goods. So you've got everything. So you've got your really high percentage distillates 178 00:27:56,280 --> 00:28:06,360 to maybe just some lower percentage waxes and other oils. And then, I mean, there are 179 00:28:06,360 --> 00:28:16,800 concentrates being made and sold to edible producers. And then how do you classify something 180 00:28:16,800 --> 00:28:24,960 like just like infused, like, you know, when they make like moon rocks or something like 181 00:28:24,960 --> 00:28:29,160 that. 182 00:28:29,160 --> 00:28:34,100 Because you know, moon rocks would probably be technically a concentrate. But that's hardly 183 00:28:34,100 --> 00:28:49,480 the same thing as oil. But taking it as it, taking what we have, taking our lemons, we'll 184 00:28:49,480 --> 00:28:58,280 try to make lemonade. We've got the sheer flour and the sheer concentrate, which are 185 00:28:58,280 --> 00:29:11,200 actually by themselves are interesting figures. So now we can calculate the cannabis CPI, which 186 00:29:11,200 --> 00:29:20,840 will be the flour share of prices times the retail flour price plus the concentrate share 187 00:29:20,840 --> 00:29:29,960 of sales times the retail concentrate price. And so this, you can maybe calculate this 188 00:29:29,960 --> 00:29:39,720 in a more glamorous way. But that is essentially this calculation here. So we'll calculate 189 00:29:39,720 --> 00:29:48,440 the CPI. Let's just look at it just to see what it may look like. However, the nominal 190 00:29:48,440 --> 00:30:00,480 value doesn't matter too much. It's more about the trend. So the trend, we can calculate 191 00:30:00,480 --> 00:30:12,600 what we've defined as inflation, which is just the rate of change of the CPI. We'll 192 00:30:12,600 --> 00:30:16,760 basically just be looking at the CPI. And so just to go back to the presentation real 193 00:30:16,760 --> 00:30:27,560 quick. So it's basically the CPI of today minus the CPI of yesterday divided by the 194 00:30:27,560 --> 00:30:35,800 CPI of yesterday. Last week, I just made a crude hypothesis that it would be between 195 00:30:35,800 --> 00:30:43,720 1 to 3% on average, which would maybe be what you'd expect then maybe in a more mature 196 00:30:43,720 --> 00:31:09,560 industry. So let's go ahead and calculate this and run it. And I'm actually, so it's 197 00:31:09,560 --> 00:31:18,920 kind of jumping across the board. The first thing I notice is it is negative, which negative 198 00:31:18,920 --> 00:31:32,880 inflation is deflation. Deflation from the federal level is not desirable. So deflation 199 00:31:32,880 --> 00:31:43,160 can have bad effects. So from a consumer's point of view, if there is deflation, then 200 00:31:43,160 --> 00:31:49,480 you know that prices are going to be decreasing. So you know prices next week are going to 201 00:31:49,480 --> 00:31:58,880 be lower than prices today. So it gives you the incentive to wait to buy your goods next 202 00:31:58,880 --> 00:32:07,360 week. So if everybody's just waiting as long as they can to buy their goods, then it's 203 00:32:07,360 --> 00:32:17,840 going to decrease the economic demand even more. So businesses, they're already struggling 204 00:32:17,840 --> 00:32:27,840 for demand. And now consumers are postponing their purchases. So the businesses don't have 205 00:32:27,840 --> 00:32:35,480 much cash flow. So it's even harder for the businesses. They lower their prices even further. 206 00:32:35,480 --> 00:32:44,640 And so you can get trapped in essentially the cyclical deflation. That's sort of bad 207 00:32:44,640 --> 00:32:50,680 in the economy as a whole. Like I said, the cannabis industry, it's sort of this short 208 00:32:50,680 --> 00:33:00,760 term shock where things needed to stabilize. There was just this, I think there was even 209 00:33:00,760 --> 00:33:05,760 a note in Oregon where they're just saying, oh, there's so much supply. So just in this 210 00:33:05,760 --> 00:33:11,240 first few years, everyone's sort of figuring out, okay, what is the level of supply? What's 211 00:33:11,240 --> 00:33:22,560 the level of demand? And prices, it appears, were just too high. Or not too high, but maybe 212 00:33:22,560 --> 00:33:29,600 those were the prices maybe when you were expecting more of a gray slash transitioning 213 00:33:29,600 --> 00:33:39,320 from a black market. And then as more producers entered the market, you see things stabilize 214 00:33:39,320 --> 00:33:50,040 over time. So it looks like it took about two years for people who were thinking about 215 00:33:50,040 --> 00:33:57,000 entering the market to enter. So this is the time essentially where everybody's entering 216 00:33:57,000 --> 00:34:05,660 the market and the people who aren't successful maybe exiting the market. And then it looks 217 00:34:05,660 --> 00:34:15,640 like by about 2019, things are starting to stabilize. So let's actually just look at 218 00:34:15,640 --> 00:34:32,060 inflation in percentages real quick, because that's typically what you think about. Then 219 00:34:32,060 --> 00:34:41,640 let's just look at the mean. So as a whole, prices are decreasing. I would just note that 220 00:34:41,640 --> 00:34:54,480 if you just look at the last two years, if you just look at the last two years, inflation 221 00:34:54,480 --> 00:35:05,560 maybe is around 0.5% per month, which is maybe that's closer to typical. So maybe moving 222 00:35:05,560 --> 00:35:16,560 forward, you may expect more, this slight 0.5% a month inflation. Maybe there's some 223 00:35:16,560 --> 00:35:30,680 shocks here and there, but personally, I wouldn't expect this deflation to continue. So that's 224 00:35:30,680 --> 00:35:40,640 just sort of our crude data analysis. So now let's grab the interest rate and start forecasting 225 00:35:40,640 --> 00:35:56,000 here. So you can get your API key for free from the St. Louis Fred. They're a good resource. 226 00:35:56,000 --> 00:36:01,640 Just starting to try to incorporate them into my work just to get some more data points 227 00:36:01,640 --> 00:36:15,760 in here. This is the effective federal funds rate. It can be a good starting point. And 228 00:36:15,760 --> 00:36:40,120 it looks like something went wrong here. Okay. So not 100% sure. I'll have to investigate 229 00:36:40,120 --> 00:36:45,680 what that was after this. But long story short, we've got the interest rate now. I think that 230 00:36:45,680 --> 00:36:54,240 was maybe some sort of error maybe reading in my API key. Anyways, moving on. We've got 231 00:36:54,240 --> 00:37:10,000 our interest rate now. So we've basically just read in the tail end here of the interest 232 00:37:10,000 --> 00:37:18,800 rate. And we'll use it in our vector autoregression. And like I said, next month we can compare 233 00:37:18,800 --> 00:37:25,520 our forecasts and perhaps even maybe even next week we can maybe even try a different 234 00:37:25,520 --> 00:37:34,520 forecasting model and then compare the two forecasting models. So we could try next week 235 00:37:34,520 --> 00:37:42,320 to just do an autoregression of inflation. And then we can compare that to the vector 236 00:37:42,320 --> 00:37:51,720 autoregression and see what forecasts better, the more complex model or the more simple 237 00:37:51,720 --> 00:38:04,240 model. And that takes us to our 10 commandments where we wanted to use recursive methods. 238 00:38:04,240 --> 00:38:20,720 Okay. So we have our total sales, we have inflation, and we have the interest rate. 239 00:38:20,720 --> 00:38:27,200 We're going to toss all of these together into a vector. And so just to show you what 240 00:38:27,200 --> 00:38:43,080 the vector looks like. So basically this is sales, this is inflation, and this is the 241 00:38:43,080 --> 00:39:00,200 interest rate in time period zero, which would be 2017 January. So this is sort of what the 242 00:39:00,200 --> 00:39:12,320 vector looks like. So just to show you what's happening under the hood. And so now I wrote 243 00:39:12,320 --> 00:39:17,920 a scrappy vector autoregression and I wasn't actually very pleased with how it was working. 244 00:39:17,920 --> 00:39:26,280 So I actually noticed that you can actually just use a vector autoregression in the stats 245 00:39:26,280 --> 00:39:37,840 models package. So it saved a lot of time. And it's a bit more flexible. You've got more 246 00:39:37,840 --> 00:39:46,160 contributors, so more stable than my scrappy var model. So we'll use this one from this 247 00:39:46,160 --> 00:40:03,680 package. And so let's say you're just going to fit. Well, this one has my notes on it. 248 00:40:03,680 --> 00:40:14,040 So here's our three equations. So let's go ahead and fit this. So output inflation, the 249 00:40:14,040 --> 00:40:29,720 interest rate. And we're just going to use one lag order. And so what we've done here 250 00:40:29,720 --> 00:40:41,640 is we've basically estimated three ordinary least squares regressions simultaneously. 251 00:40:41,640 --> 00:40:54,680 So this is our first regression. This is our regression for output. Then this is our regression 252 00:40:54,680 --> 00:41:07,800 for inflation. And then this is our regression for the interest rate. And so as you see, 253 00:41:07,800 --> 00:41:18,120 each one has a constant. And then these are the coefficients. So this is beta, this is 254 00:41:18,120 --> 00:41:40,360 gamma, and this is delta. So we've estimated quite a lot of parameters here. I'm not sure 255 00:41:40,360 --> 00:41:52,320 if they tell us our degrees of freedom here. But long story short, we were able to fit 256 00:41:52,320 --> 00:42:03,980 an AR1. So that's a good sign. Next. If you were going to fit multiple models, so say 257 00:42:03,980 --> 00:42:20,600 we're going to fit an AR2. So we're now going to use two lags. So here are our three simultaneous 258 00:42:20,600 --> 00:42:36,200 equations. And now you'll notice that we have two lags of each. Yes, so two lags of each 259 00:42:36,200 --> 00:42:45,560 variable. So there's output lagged once, output lagged twice. Inflation lagged once, inflation 260 00:42:45,560 --> 00:42:54,200 lagged twice, interest rate lagged once, interest rate lagged twice in all three models. So 261 00:42:54,200 --> 00:43:05,400 that is a lot of parameters. That is 21 parameters, I believe. So that's one, right? So that's 262 00:43:05,400 --> 00:43:16,920 six, seven times three should be 21 parameters. So I'm not sure if we can, it's going to be 263 00:43:16,920 --> 00:43:26,040 tough to estimate too much larger. However, we can try. But we need some way to compare 264 00:43:26,040 --> 00:43:32,240 these models to each other, right? Because we can just do this all day if we have a large 265 00:43:32,240 --> 00:43:45,240 enough data series. So centrally, we need to have a criterion. Centrally, the BIC, the 266 00:43:45,240 --> 00:43:55,400 Bayesian Information Criterion, it rewards you for making better predictions, but then 267 00:43:55,400 --> 00:44:05,680 it punishes you for adding more and more predictors, more and more parameters. So the BIC is a 268 00:44:05,680 --> 00:44:16,160 useful way to measure which model is a better, which model may be a better choice. So as 269 00:44:16,160 --> 00:44:23,000 you'll notice, this model, we tossed in a ton of parameters, so we may be able to predict 270 00:44:23,000 --> 00:44:33,200 a little better. But we've got a BIC of 22, and this model has a BIC of 21. And so we 271 00:44:33,200 --> 00:44:46,600 want to pick the model with the minimum BIC. And so in this situation, the AR1, it's more 272 00:44:46,600 --> 00:44:57,400 parsimonious and it's actually the better model choice if you're choosing by BIC, because 273 00:44:57,400 --> 00:45:08,560 you're essentially overfitting with this model. And it may lead to bad out of sample predictions. 274 00:45:08,560 --> 00:45:17,160 So that's sort of a long explanation of what we're about to let stats models do for us 275 00:45:17,160 --> 00:45:25,440 automatically. So we can actually just let stats models say, OK, why don't you try every 276 00:45:25,440 --> 00:45:38,080 model up to lag order six and just tell me the best one based on minimum BIC. So we can 277 00:45:38,080 --> 00:45:44,080 just let stats models do that for us and save us a lot of programming, because you could 278 00:45:44,080 --> 00:45:53,280 program that up yourself. But you let stats models do it. And look at that. They tried 279 00:45:53,280 --> 00:46:01,440 every model that they could fit up to lag order six. I'm sure the higher order lags, 280 00:46:01,440 --> 00:46:08,160 they weren't able to fit successfully. And so they determined, OK, the model with the 281 00:46:08,160 --> 00:46:18,640 minimum BIC is actually the lag order one. So that's just our first model. 282 00:46:18,640 --> 00:46:28,480 What this is is this model, but without the dot, dot, dots. So we're literally just doing 283 00:46:28,480 --> 00:46:40,840 YT on YT minus one, inflation of T minus one, the interest rate of T minus one, and so forth. 284 00:46:40,840 --> 00:46:52,240 So we finally get to the goods. So we've got our best model. It's forecasted and show the 285 00:46:52,240 --> 00:47:16,520 data. Maybe I ran that twice. Not certain why it's printing twice. But anywho, here 286 00:47:16,520 --> 00:47:27,600 are our forecasts. So this legend's a little in the way, but there's our forecast for our 287 00:47:27,600 --> 00:47:44,960 total sales. Here is our forecast for inflation. So slightly rising. And then our, see, this 288 00:47:44,960 --> 00:47:55,160 is interesting. So our forecast for the interest rate is negative. But as we know, the interest 289 00:47:55,160 --> 00:48:04,880 rate can't be, well, it can't be nominally below zero. So the Federal Reserve has to 290 00:48:04,880 --> 00:48:19,240 use monetary policy, quantitative easing, and other tools in their toolbox to effectively 291 00:48:19,240 --> 00:48:31,620 drive the interest rates below zero, if that's with their policy. So that's the plot. I will 292 00:48:31,620 --> 00:48:41,060 admit that I am just now learning the stats models packages, because I would like to actually 293 00:48:41,060 --> 00:48:55,720 print out with this series is here. But I am not certain that I know the thing. I may 294 00:48:55,720 --> 00:49:03,020 need to learn the stats models package a little bit better to actually get you these raw data 295 00:49:03,020 --> 00:49:17,340 points here. However, here is the package I'm using. And there is a lot here. So I've 296 00:49:17,340 --> 00:49:30,220 just hit the surface. You can do error analysis to make sure you're not doing anything crazy. 297 00:49:30,220 --> 00:49:40,680 We've done the lag order selection. And then I think for maybe in the future, I would like 298 00:49:40,680 --> 00:49:48,300 to show you how to do impulse response function. Because you can actually say, okay, well, 299 00:49:48,300 --> 00:49:56,500 what would actually happen if there was a shock in prices? How would that affect output? 300 00:49:56,500 --> 00:50:06,700 Or if there was a shock in the interest rate, how would that affect prices or sales? So 301 00:50:06,700 --> 00:50:14,460 all three of these series are theoretically related to each other, particularly sales 302 00:50:14,460 --> 00:50:28,460 and prices. So long story short is you can use these impulse response functions to simulate 303 00:50:28,460 --> 00:50:35,780 shocks. So what would happen if all of a sudden there was a recession? So these could be useful 304 00:50:35,780 --> 00:50:47,740 for simulating harvest spikes in October or something like that. So perhaps for next week 305 00:50:47,740 --> 00:51:00,540 we can dive into that. And I'll actually want to show you the actual numbers here. But that 306 00:51:00,540 --> 00:51:11,700 is sort of the conclusion of my spiel here on forecasting inflation. And so I just maybe 307 00:51:11,700 --> 00:51:20,580 like to open it up if there are any questions or anything. 308 00:51:20,580 --> 00:51:37,540 Would be able to, is that beneficial or there are avenues for better analysis next week? 309 00:51:37,540 --> 00:51:38,540 Or? 310 00:51:38,540 --> 00:51:46,180 No, it was interesting. These are things that I don't think about. So it was good to learn 311 00:51:46,180 --> 00:51:57,940 about how you go about forecasting these things. So it was cool. I learned something. 312 00:51:57,940 --> 00:52:04,420 What are some of the things you think about? Because this may have been, I'm not sure, 313 00:52:04,420 --> 00:52:09,620 this may have been a bit too economics heavy. So what are some of the things that you're 314 00:52:09,620 --> 00:52:13,140 thinking about on a day to day basis? 315 00:52:13,140 --> 00:52:19,540 Well, that's kind of in my data science journey. I realized there's like a data science food 316 00:52:19,540 --> 00:52:24,620 pyramid of things that you should know. And so I know everything at the top and I know 317 00:52:24,620 --> 00:52:31,620 everything at the bottom, but there's this middle business understanding, business storytelling 318 00:52:31,620 --> 00:52:39,220 kind of thing, which I don't know. And so I'm trying to learn that kind of thing. And 319 00:52:39,220 --> 00:52:43,260 these presentations have been really good for that. 320 00:52:43,260 --> 00:52:55,740 Well, that could also just be my style. So that's sort of, from my experience, that was, 321 00:52:55,740 --> 00:53:01,820 so I attended a lot of seminars when I was in college. And that's what I noticed made 322 00:53:01,820 --> 00:53:11,100 the most successful presentations slash papers is when you sort of tell a story and you let 323 00:53:11,100 --> 00:53:20,180 the models evolve. So you start with some sort of business question. So here we started, 324 00:53:20,180 --> 00:53:27,060 it helps to start with just a figure. So start with the data. So we just started with what 325 00:53:27,060 --> 00:53:35,100 we're looking at sales in Oregon, prices in Oregon. And we sort of just let the model 326 00:53:35,100 --> 00:53:44,660 evolve naturally. So we start as simple as we can with essentially the growth rate. So 327 00:53:44,660 --> 00:53:53,580 you start with the simplest statistics you can calculate, and then you gradually make 328 00:53:53,580 --> 00:54:00,860 it more complex. So really before you jump to the VAR, the VAR, you'd really want to 329 00:54:00,860 --> 00:54:07,580 just estimate a VAR, which we've kind of done. Well, I'm actually, I'm not sure, I'll have 330 00:54:07,580 --> 00:54:16,940 to review it. We may have done that prior weeks, but essentially you let the model become 331 00:54:16,940 --> 00:54:31,540 gradually more complex, but still keeping in mind your initial objective. And then you 332 00:54:31,540 --> 00:54:48,580 just, yeah, you, and then as you're doing your analysis, you'll be calculating other 333 00:54:48,580 --> 00:54:55,020 metrics of interest. So we calculated the share of flour, we calculated the share of 334 00:54:55,020 --> 00:55:07,020 concentrates. All of those sort of add to the story we're telling here of this presentation. 335 00:55:07,020 --> 00:55:20,860 So you sort of build these pieces here. They're just sort of pieces of your analysis that 336 00:55:20,860 --> 00:55:28,180 just gradually gets, you just keep adding to it until you've, you know, you've done 337 00:55:28,180 --> 00:55:35,900 sort of an in-depth analysis on the subject at hand. And then you've answered sort of 338 00:55:35,900 --> 00:55:44,740 your question at hand, which started by just looking at the data. We just basically wanted 339 00:55:44,740 --> 00:55:58,500 to know what's the trajectory for inflation. What even is inflation? So we were able to, 340 00:55:58,500 --> 00:56:06,980 you know, gradually build up to that. And then we can answer that question. We can say, 341 00:56:06,980 --> 00:56:15,820 oh, it looks like despite, you know, historic deflation in the first two years, you know, 342 00:56:15,820 --> 00:56:20,420 we predict, yeah, moderate inflation. And I wish I could print the numbers and actually 343 00:56:20,420 --> 00:56:29,900 give you the number here. But yeah, we predict, you know, moderate inflation here throughout 344 00:56:29,900 --> 00:56:39,460 the rest of the year. And you could keep taking this story further and keep adding steps to 345 00:56:39,460 --> 00:56:45,700 it. So you could say, oh, like, what's going to happen if there's a harvest spike or what 346 00:56:45,700 --> 00:56:55,020 about this? Or can't we estimate an even better model? So, and that's where sort of papers 347 00:56:55,020 --> 00:57:00,740 build on each other is because then you kind of can continue someone else's story by saying, 348 00:57:00,740 --> 00:57:09,060 oh, picking up this story where they left off, I'm going to, you know, add to their 349 00:57:09,060 --> 00:57:20,140 model. But that's my approach is just start off simple and then just sort of document 350 00:57:20,140 --> 00:57:24,460 your steps. You know, instead of just jumping to the end and just saying, oh, I'm just going 351 00:57:24,460 --> 00:57:31,660 to do a, you know, principal components analysis, or I'm going to do, you know, machine learning 352 00:57:31,660 --> 00:57:39,520 algorithm, XYZ, you know, instead of just jumping all the way there, I find it's more 353 00:57:39,520 --> 00:57:48,300 informative to start with the basics and then build your way up until the machine learning 354 00:57:48,300 --> 00:57:58,740 algorithm is actually like the next reasonable next step. Because eventually, if you kept 355 00:57:58,740 --> 00:58:04,000 taking this analysis further and further and further, I mean, machine learning would be 356 00:58:04,000 --> 00:58:10,700 the rational next step. I'm just not, I just wouldn't start with it right out of the gates. 357 00:58:10,700 --> 00:58:20,620 So that's a long-winded answer, but, you know, I resonate with that statement. 358 00:58:20,620 --> 00:58:28,020 Yeah, cool. That's, yeah, that's, you know, I think kind of the lesson that I've been 359 00:58:28,020 --> 00:58:32,660 learning and everybody emphasizes machine learning a lot, but actually there's a lot 360 00:58:32,660 --> 00:58:44,060 of steps before you get there and before it's even useful. Recruiter. They're like, they 361 00:58:44,060 --> 00:58:50,460 never stop calling. But anyway, I read a really cool article in the Oregonian and I'm trying 362 00:58:50,460 --> 00:58:53,940 to figure out how to share it because the Oregonian is making it more and more difficult 363 00:58:53,940 --> 00:59:01,420 to read their articles online and share them. But there was an article about the counties 364 00:59:01,420 --> 00:59:10,860 and the towns along the Oregon-Idaho border and how, because, you know, cannabis is legal 365 00:59:10,860 --> 00:59:16,420 in Oregon, but not in Idaho, that these towns have, you know, like on the weekend, they 366 00:59:16,420 --> 00:59:28,160 have huge, huge influxes of people from Idaho coming in and like the per capita cannabis 367 00:59:28,160 --> 00:59:33,580 sales in these towns is like four and five times the amount it is in like Multnomah County, 368 00:59:33,580 --> 00:59:38,660 which is like the most populous county in Oregon. Oh gosh. Oh, because people are coming 369 00:59:38,660 --> 00:59:42,900 in from Idaho. Coming in, yes. And so, and, you know, and there was a statement from like 370 00:59:42,900 --> 00:59:48,140 some, somebody in New Mexico and they're like, yeah, we needed to legalize it before Texas 371 00:59:48,140 --> 00:59:56,820 does so that we can, you know, we can capitalize on those sales. So it was really interesting. 372 00:59:56,820 --> 01:00:03,780 I guess that's one of those things where it's so hard to capture that in the data. And I'm 373 01:00:03,780 --> 01:00:12,580 not even sure what the laws on that are. Maybe, but that's up to the individuals who are doing 374 01:00:12,580 --> 01:00:20,180 that, I suppose. But I know there's a town called Ontario, Oregon, that's right on the 375 01:00:20,180 --> 01:00:28,420 Oregon, Idaho border that has a pretty big dispensary. So, well, has a dispensary. So 376 01:00:28,420 --> 01:00:32,980 they're making money from both states. Well, the people in Oregon are just saying, hey, 377 01:00:32,980 --> 01:00:41,300 we're just, we're just opening up shop. We're not doing anything wrong. Yeah. I don't know. 378 01:00:41,300 --> 01:00:48,900 I don't know. That's a, that's an interesting observation. And I, yeah, that's something 379 01:00:48,900 --> 01:00:53,940 to think about Charles, if there's any way to parse that out of the data. I don't know 380 01:00:53,940 --> 01:01:08,220 if that Oregon data has a geographical, like component to it, but I don't know. That may 381 01:01:08,220 --> 01:01:13,720 take a bit more exploration. I'm not sure if you can get sales by licensee in Oregon 382 01:01:13,720 --> 01:01:20,100 or not and try to do a geographical analysis. Well, I mean, whoever wrote, whoever wrote 383 01:01:20,100 --> 01:01:25,340 this article is able to figure it out. They were able to do like a per capita analysis. 384 01:01:25,340 --> 01:01:32,860 Oh, okay. Okay. So you may, they may have county by county data. So you may actually 385 01:01:32,860 --> 01:01:40,980 be able to do some sort of analysis there. So that perhaps actually for a future meetup, 386 01:01:40,980 --> 01:01:46,980 we could do geographical analysis there because like I said, I think the geographical dimension 387 01:01:46,980 --> 01:01:53,220 is an incredibly interesting one. It's can be tricky, but we've got the counties. So 388 01:01:53,220 --> 01:02:00,740 that, I mean, there's no reason why we can't do some sort of analysis. Yeah. Yeah. And 389 01:02:00,740 --> 01:02:06,740 that Ontario is only like Oregon's only like less than an hour away from Boise, Idaho. 390 01:02:06,740 --> 01:02:12,020 So I've met Boise's people from Boise are like, cause Boise is a college town too. So 391 01:02:12,020 --> 01:02:21,620 you, you probably have a lot of people from Boise just drive down to here to pick up stuff. 392 01:02:21,620 --> 01:02:25,940 That would be a real interesting analysis. And you could extend that to a bunch of states 393 01:02:25,940 --> 01:02:33,020 because that would be interesting to see if you have like this like border phenomena in 394 01:02:33,020 --> 01:02:39,780 cannabis states where of course you would expect like the cities to be like high in 395 01:02:39,780 --> 01:02:47,260 sales, but then it would be so funny if like the border towns were all high sales too. 396 01:02:47,260 --> 01:03:01,180 Yeah. That's interesting observation and could be a good opportunity for analysis. So I think 397 01:03:01,180 --> 01:03:09,540 I'll go ahead and just, unless anyone's got any more topics at hand to talk about it, 398 01:03:09,540 --> 01:03:19,020 we've kind of reached the hour. So let's anyone else have to throw on the table? I mean, go 399 01:03:19,020 --> 01:03:29,180 ahead and go ahead and conclude it here until next week. All right. All right. Well, thank 400 01:03:29,180 --> 01:03:35,420 you all for coming. Thanks, Ryan. Thanks Charles. Thanks Nick. Thank you for your contributions. 401 01:03:35,420 --> 01:03:41,140 It's always, it's always awesome to hear from you. And then next week we'll see if we can't 402 01:03:41,140 --> 01:03:51,140 look at Charles's waste data, touch up the inflation data if needed. And then, and then 403 01:03:51,140 --> 01:04:02,820 maybe start some geographical analysis of if we can find that county data. All right, 404 01:04:02,820 --> 01:04:10,980 everyone. Thank you. Thanks. Thanks for coming. And until next week, have an awesome week. 405 01:04:10,980 --> 01:04:22,740 Bye. Bye. Bye.